This PDF file contains the front matter associated with SPIE Proceedings Volume 13213, including the Title Page, Copyright information, Table of Contents, and Conference Committee information.

This article proposes a new collection of orthogonal image representations known as Generating Function Moments (GFMs). The functions related to the image kernel of GFMs originate from the generating function system, a set of orthogonal spline basis functions with a unit denoted as k, allowing for values of 0, 1, 2, and beyond. The utilization of GFMs is not necessitate any form of numerical estimation and maintains a comparatively reduced complexity in computation due to the advantages inherent in their lower-order basis set. These attributes make, GFMs a more effective alternative to conventional moments in polynomial like Zernike and Legendre moments, particularly concerning image reconstruction. The experiments presented in this work also demonstrate that GFMs possess an enhanced capability for representing higher features.

VR technology is a computer simulation technology that can build and experience a virtual world. It uses computers to excite an interactive 3D dynamic scene, and the simulation system of its physical behavior can immerse people in the environment. Flat images are prone to color defects under the influence of light and angle, but due to the poor adaptive effect of traditional methods designed flat visual image color adjustment systems, the color differences in the adjusted images are not significant. In order to enhance the visual effect of flat image color, this article designs a flat image color enhancement system based on VR technology. Perform wavelet denoising preprocessing on the collected flat images, combine VR imaging technology for image information enhancement, identify image color defects using machine vision technology, and design an image color adjustment module using the seven color algorithm to output high-quality flat images. The experimental results show that by adjusting the image color through the designed system, the plane visual brightness and clarity of human-computer interaction (HCI) can be enhanced, and the saliency of page details can be enhanced; Compared with traditional image color adjustment systems, the designed system's color adjustment effect is more in line with expectations and meets the color requirements of the original image.

With the continuous development of TOF imaging technology, distance imaging technology has gradually been widely used as a part of three-dimensional spatial vision in many fields such as autonomous driving, mechanical obstacle avoidance, and virtual reality. However, due to the limited performance of current distance imaging sensors, the quality of depth images obtained from distance imaging is generally low, making it difficult to apply to computer vision tasks with high accuracy requirements. For the current problems of low accuracy, fewer boundary data, and poor dimensional resolution of TOF image, we introduce a TOF-CMOS image fusion approach found on DTCWT (Dual-Tree Complex Wavelet Transform). By utilizing the similarity between standard CMOS images and TOF images of the same scene in the edge area, the detailed information of the TOF images is enriched and their accuracy is improved. First, we applied the dual-tree complex wavelet transform applied to the source images. Second, we adopt a bilateral filter for the low frequency components of the TOF images to increase the accuracy of the images. Meanwhile, we extract the edge data in the high-frequency components of the CMOS images through the soft threshold method. Then, we blend the high-frequency components of the TOF images with the extracted high-frequency components of the CMOS images. In this paper, we innovatively propose fusion rules connecting the regional variance and the pixel maximum: we use the above two fusion methods respectively targeted according to the difference of data involved in the highest as well as the second-highest levels of high-frequency components. Finally, the images are reassembled by using the inverse dual-tree complex wavelet transform. We evaluate the merge results using evaluation indicators like average gradient, peak signal to-noise ratio, and so on. The laboratory results indicate that the method adequately enhances the visual effect of the TOF image and enriches the detailed information of the TOF image.

Retinal layer thickness is one of the important indicators for diagnosing ophthalmic diseases, therefore, automatic segmentation and quantitative calculation of the retinal layer are of great significance. However, currently in retinal optical coherence tomography images, there are many problems with retinal stratification, such as the lack of contextual semantic information due to blurred boundaries of each layer of the retina, and the discontinuity in the pathological change region. In this paper, a novel dense encoder and position attention method which is introduced into the encoding and decoding architecture is proposed to improve the precision of retinal layer segmentation and measurement. The dense encoder block can learn and extract more diverse the global context, and the position attention is to ensure the spatial consistency of retinal targets boundaries and regions, to improve the imbalance between retinal layer boundaries and obtain more continuous prediction classes for the same retinal layer structure. The validation results demonstrate that our proposed method achieved an average dice of 87.35%, and the average thickness error of the retinal layers was reduced to 1.74 pixels. In particular, it provides an effective strategy to improve the boundaries of segmentation and accurately segment and measure the retinal layers.

Although attention mechanisms have been widely applied in natural language processing (NLP) tasks, there are still limitations in their utilization within the field of computer vision. To integrate the advantages of convolutional neural networks (CNNs) with attention mechanisms, this study proposes a multimodal image generation model based on the Stable Diffusion architecture. The model incorporates two types of convolutional modules, namely ICB and TCB. By replacing the linear networks with convolutional neural networks, the model enhances its capability to process complex features. Subsequently, the low-dimensional encoding reconstruction of images is achieved by maximizing the mutual information between the input of the new model and the output of the encoder. Finally, the proposed model is validated using a publicly available dataset.

In this paper, we introduce an image restoration network that is based on the High-Efficiency Transformer (HET). The model utilizes fast Fourier convolution for image enhancement, incorporates self-attention mechanisms for capturing contextual information, and employs convolution for feature extraction and classification. Additionally, we improve the network's performance and convergence speed by implementing batch normalization and residual concatenation. Experimental results show substantial enhancements in image restoration achieved through our deep learning-based approach. The reconstructed images demonstrate improved clarity, enhanced details, and reduced dependence on manual interventions.

This study proposes an end-to-end method aimed at achieving intelligent colorization of images. This approach enables automatic colorization of grayscale images by taking in a grayscale image and a colored reference image as input. We designed a deep neural network model to learn the mapping relationship between grayscale images and colored reference images in an end-to-end manner. During training, we utilized a large amount of image data with real color labels and optimized the model using generative adversarial networks. Experimental results demonstrate the superiority of our method in color restoration and detail preservation across different types of images. Compared to traditional methods, the end-to-end approach offers higher levels of automation and faster colorization speed, opening up new possibilities in the field of image colorization.

U-Net has become an indispensable component in medical image segmentation tasks. The characteristic of U-Net is that it produces multi-scale features, multi-scale features can provide hidden features under different views, which helps improve semantic segmentation performance. In addition, knowledge distillation, e.g., feature distillation or logit distillation, is a mechanism that can efficiently compress models. Feature distillation guides students’ feature learning by transferring feature information. In order to be able to supervise and distill these multi-scale features in feature distillation, we propose a Multi-scale Feature Distillation (MFD). MFD uses the teacher's predicted logits as the distillation target, and the students' multi-scale features of different layer as the supervision target. Nowadays, it has become a trend to decouple logits distillation. Original logits distillation can usually be divided into target classes and non-target classes. Target classes and non-target classes often play different roles in feature distillation and logits distillation. We introduce a Decoupled Multi-scale Distillation (DMD) that utilize target classes and non-target classes for feature distillation and logits distillation. When performing feature distillation, we use non-target classes for distillation, and when performing logits distillation we use target classes for distillation. Experiments on different datasets demonstrate that the DMD is effective.

This thesis delves into the significance and challenges of Near-Duplicate Image Detection (NDD) in today's digital media landscape, where the varied manipulation of images complicates accurate identification. By designing and implementing a Near-Duplicate Image Detection method based on the Scale-Invariant Feature Transform (SIFT) algorithm, the paper aims to assist us in quickly and accurately identifying possible image duplication, cropping or other attacks. By detailing the key steps of the SIFT algorithm, including feature point extraction, outlier descriptor rejection strategy, and efficient matching using the Approximate Nearest Neighbor (Annoy) algorithm, the paper demonstrates the advantages of the method in improving detection efficiency and accuracy. In addition, the potential of image feature extraction in dealing with high-resolution image alignment is emphasized, as well as its importance for approximate duplicate detection and image copyright awareness protection. Through comparative experimental validation, the paper demonstrates that approximate image detection algorithms based on SIFT and Annoy algorithm have significant advantages in the field of image matching for large datasets, providing an effective tool for image detection and academic research.

To effectively and accurately detect regions with subtle changes in wide-field videos captured by eagle-eye equipment in low-light environments. A two-level change detection (CD) algorithm based on guided filter fusion is proposed to achieve better separation in complicated scenarios. First, the difference images (DIs) is obtained through the difference operator and the extreme minimum ratio operator. At the same time, using Principal Component Analysis (PCA) fuse the global brightness component and the V component are extracted from images to obtain the optimized brightness map. Then, the saliency map is obtained from the optimized brightness map based on the maximum salient pixel and spatial continuity characteristics as a guide for DIs fusion. Finally, a two-level method is proposed to combine the threshold method and clustering algorithm to further suppress the noise and DI normalization. Then, using the K-means algorithm to to generate the final binary change map. Experimental results show that the proposed algorithm can accurately and effectively detect changes in Eagle Eye surveillance video images in low-light environments. Compared with other algorithms, this algorithm reduces the running time and effectively minimizes the interference of light sources and dark spot noise in low-light environments.

With the development of generative adversarial networks (GANs) and diffusion models, deepfake has become a serious security issue. This study mainly aims to detect images produced by diffusion models. For this reason, this paper introduces an innovative Local Frequency Analysis (LFA) method which based on Krawchouk moment mesoscale frequency analysis and Discrete Cosine Transform (DCT) finescale frequency analysis, which simultaneously performs the frequency domain analysis of images from multi-scale to extract more comprehensive features. In this paper, experiments are performed on two datasets, GenImage and DiffusionForensics, and the proposed method is evaluated and contrasted against the existing methods. Also, the generalization capability of LFA for detecting forged images from different sources is experimented on the GenImage dataset. The results of the experiments show that LFA performs superiorly in detecting diffusion-generated images. In addition, robustness experiments on two types of degradation, JPEG compression and Gaussian blur, are also conducted to further demonstrate the practicality of LFA.

The method proposed in this study addresses the issues of existing infrared imaging devices on the market, such as limited detection of target objects, large size, narrow detection range, and lack of mobility. It presents a solution for infrared thermal imaging to locate and track anomalous temperatures on a mobile platform. This method is capable of detecting abnormal facial information within a certain range, determining facial positions. In practical scenarios, the horizontal position of anomaly information cannot be accurately obtained due to possible occlusions, thus necessitating spatial localization, which involves calculating the horizontal position and angular offset of the camera plane. This paper introduces a method for calculating the horizontal position of a binocular depth camera, as well as the transformation relationship between the platform pose, thermal imaging module, and the object's position in the camera, collectively determining the offset angle for tracking purposes. Additionally, a loss function for temperature with respect to distance is fitted, enabling accurate calculation of the actual temperature of anomalies.

The smooth warp field is becoming increasingly important in image stitching. When the 3D scene is confronted, the relative positions of the scene structures observed at different camera views may differ, thus causing parallax problems. Existing image stitching algorithms often fail in parallax scenes and are not efficient enough to meet the requirements of real-time tasks such as panoramic stitching for UAVs. In this paper, we propose a deep multilayer image stitching framework for parallax scenes based on GPU parallel computing, which includes multi-scale RANSAC layers, a parallax suppression layer and a smooth projection field generation layer. Instead of strictly dividing the RANSAC process and the dense warping model into two steps as in traditional image warping, we use the dense warping model for outlier filtering in the RANSAC process, and the two steps are performed in a cross iterative fashion. Experimental results show that our algorithm significantly outperforms the existing algorithms in terms of computational performance and achieves an approximate level of quality of stitching.

Addressing the problem of poor visual effect and noise interference in foggy video images collected by hawk-eye image sensors, we suggest an approach which utilizes difference images (DIs) fusion and two-stage clustering for detecting changes in monitoring video pictures. First, we utilize a dehazing algorithm based on the color attenuation prior theory to effectively eliminate the haze interference. Then, two DIs are generated using the compressed logarithmic ratio and maximum mean pixel ratio operators. Fusion of two DIs using multi-scale Laplace pyramid and regional energy algorithms, and filtered using the adaptive median filtering algorithm to obtain the final DIs. Finally, the soft threshold function and the FCM clustering algorithm are used to perform two-stage clustering on the DI to obtain the change detection result. Experimental results demonstrate that this approach can effectively and accurately detect subtle changes in hawk-eye surveillance video images under foggy conditions.

Video surveillance requires simultaneous monitoring of multiple areas. Consequently, real-time automatic change detection of the monitored areas becomes very important. In the context of wide field-of-view conditions, the combination of a wide field-of-view, intricate environmental factors, and a substantial presence of random noise can lead to the degradation of visual fidelity and a diminished signal-to-noise ratio in the video images acquired through the image sensor. As a consequence, the task of detecting subtle changes becomes challenging for the surveillance system. To address the above problems, we have proposed a change detection method that leverages improved difference images and super fast and robust fuzzy c-means with constraints clustering. Initially, we employ an improved log-ratio operator and an improved mean-ratio operator to generate two distinct difference images. Subsequently, the wavelet fusion algorithm is applied to merge these two difference images, effectively integrating their distinctive features and producing a fused difference image with differentiability. Then, the new difference image is subjected to soft threshold primary classification and a cumulative distribution function normalization to obtain the difference image after primary classification. Finally, the super fast and robust fuzzy c-means with constraints clustering algorithm is employed for the ultimate classification, enabling the separation of changed and unchanged areas within the image.

Image segmentation remains a challenging problem in computer vision. Various segmentation methods have been developed, including traditional methods based on threshold, edge, region, and morphology, as well as novel methods based on deep learning. The segmentation precision is continuously improving, and its applications are expanding. Starting with traditional image segmentation methods, we organized their performance and practical application in relation to the demand for segmentation of IC surface defects. We then tested the segmentation of the IC surface defects database we created. However, the final result was unsatisfactory. Therefore, we are continuing to explore segmentation models based on deep learning and summarise their performance evaluation to explore the possibility of applying them to the segmentation of IC surface defects.

Rice occupies a significant proportion of the national food security system, and weeds are an important factor affecting rice yield. In the natural environment, the collection of rice field images will be due to the weather, light intensity changes and other factors, resulting in the details of the image are not obvious enough, thus making the image segmentation accuracy is not high. This paper proposes an adaptive histogram equalization method for this problem, which automatically adjusts the parameters according to the luminance distribution of the local region to improve the contrast of the image and retain the image details. In order to improve the segmentation accuracy, this paper proposes Laplace pyramid image fusion, which fuses the original image with the improved image, and is able to naturally and smoothly fuse the two images while preserving the image details. In order to suppress noise and avoid single features being affected by noise, which leads to inaccurate segmentation results, this paper proposes to use the method of fusion of color features and morphological features to enhance the segmentation accuracy of rice and weeds. The experimental results show that the Accuracy of the image segmentation method proposed in this paper reaches 96.9%, Precision reaches 93.3%, Recall reaches 95.8% and MIoU reaches 86.5%.

Aiming at the issue of single-image super-resolution methods relying solely on the input images of poor resolution, which struggle to reconstruct satisfactory results when the details in the low-resolution images are insufficient, this paper proposes a text-guided images of low quality. The shortcomings of inadequate information in images with low resolution are compensated for by introducing text information as prior knowledge into the network. Specifically, a multi-modal network structure is proposed, which utilizes a text-image cross-modal pre-trained large model to extract shallow image feature information of different modalities. After deep feature extraction, sub-pixel convolution is employed for magnifying the images and reconstruct high-quality images. Experimental results indicate that the proposed method can utilize text information to reconstruct abundant details missed in low-resolution image scenarios. NIQE and PI metrics confirm the superior super-resolution performance of the approach.

A multi-level feature weighted fusion based image classification method for silicon single crystal growth process has been proposed to address the issue of the long silicon single crystal pulling times at industrial sites, which prevents timely understanding of the crystal growth process. A multi-level feature weighted fusion method was employed to identify and classify the growth process of silicon single crystals. A Czochralski single crystal silicon growth process recognition system was then designed. A comparison of the experimental results obtained from this design with the traditional single image classification method clearly showed that using different levels of silicon single crystal image information simultaneously effectively improved the accuracy of silicon single crystal growth image classification.

In the domain of remote sensing, Change Detection (CD) from bi-temporal images is crucial. Images from Hawk-Eye platforms provide an adjunct to traditional remote sensing data, but introduce significant challenges, especially under low-illumination and wide field-of-view conditions. These issues include complex noise and minute change regions, making the direct application of conventional CD techniques largely ineffective. To address this, we propose a dedicated plug-and-play module for CD on Hawk-Eye platforms. This innovative module integrates a Dark Channel Enhancement Module (DCE) to improve brightness with a Scale-decoupled Fine-slicing Module (SF) for detailed change region detection. Thorough evaluations on real images from Hawk-Eye platforms demonstrate the module's effectiveness in adapting existing CD methods.

This paper proposes a new neural network model, series connection UNet (SC-UNet). The network model is composed of two series of UNet modules. The residual module and the attention mechanism gate are added to the model to make the back-propagation faster and focus the model's attention on the target region. In the medical image segmentation task of lung X-ray pictures, compared with the single UNet model, the average crossover ratio, mean pixel accuracy, average accuracy rate, and mean recall rate of the model structure increased by 4.51%, 3.98%, 0.45%, and 4.00% respectively. Experiments show that the SC-UNet model has significant advantages over the UNet model in medical image segmentation.

Brain tumor is a disease characterized by abnormal cell proliferation occurring within brain tissue or beneath the meninges, which may pose a serious threat to the health of patients. In medicine, the diagnosis of brain tumors is mainly based on magnetic resonance imaging (MRI) information. However, manually detecting lesions in brain tumor images is a tedious, time-consuming, and error prone task. Therefore, there is an urgent need for computer-aided methods with higher accuracy and efficiency to analyze brain tumors. Deep learning can quickly extract deep features of images and is widely used in the field of medical image classification and recognition. In this study, 7023 open source MRI images of brain tumors were obtained from the Kaggle database, including four types: glioma, meningioma, pituitary gland, and tumor free. Subsequently, a deep DenseNet network framework was introduced to construct a classification model. The results show that the DenseNet121 deep learning method has a model evaluation and validation accuracy of 98.32% on the test set. Finally, the visualization analysis of gradient weighted class activation maps provides interpretability for the detection results of deep learning models.

Artificial intelligence technologies represented by deep learning have been widely applied in remote sensing image interpretation. RSIC is an essential part of remote sensing image interpretation and one of the research hotspots in the field. RSIC involves multimodal data processing such as text and images, and can be widely applied in multiple professional fields such as remote sensing image retrieval, scene understanding and classification, intelligence analysis and rapid generation. Based on this, this article summarizes the research progress of RSIC, reviews the research progress of RSIC, and first introduces the background of RSIC. Secondly, an in-depth analysis of commonly used methods for RSIC is conducted, with a focus on four aspects: RSIC methods, comparison of method experimental results, image caption datasets, and evaluation indicators. Finally, a capability analysis of RSIC methods will be conducted, and current research achievements will be further summarized, with prospects for future directions and challenges in this field.

Industrial Internet-of-Things (IoT) has recently emerged and is undergoing rapid development, and image data collected through IoT nodes or terminals is related to personal information. For the protection of image data in IoT, especially in smart grid, we propose a new lightweight image encryption and collection scheme. First of all, a random permutation matrix was used to disturb the association between original and encrypted images. Secondly, to further strengthen the encryption, we select only a portion of bits from the blinding matrix for encryption, instead of encrypting the entire image. After receiving the encrypted data, the data center performs the inverse operations to decrypt the image. Performance analysis shows that, compared to the counterparts, our scheme is more efficient and feasible, can be applied for resource-limited equipment and is ideal for protecting sensitive images.

Mural restoration has an extremely important significance in image restoration. Image enhancement aims to improve image quality, clarity or visualization. The pulse-coupled neural network (PCNN) is a single-layer neural network structure for artificial neural networks, with featuring nonlinear coupling modulation, synchronized pulses, and dynamic pulse excitation. The unique structure and working principle of PCNN enable it to perform well with strong spatial and temporal correlations in image enhancement aspect. The synaptic-linked-FCMSPCNN(SLFC-MSPCNN) is proposed in this paper, and achieves more effective control of neuron firing time by adjusting adaptive parameters. Through related experiments, The proposed SLFC-MSPCNN has good image enhancement performances comparing with popular previous image enhancement approaches in Dunhuang Murals.

Eckhorn obtained a model of a mammalian neuron, the pulse-coupled neural network model (PCNN), by studying neurons in the cat's visual cortex by examining their synchronized pulse oscillations. PCNN is a single-layer neural network model. The dynamic threshold of the model is adjustable, and it has the characteristics of nonlinear modulation coupling, synchronous pulse and dynamic pulse excitation, which makes PCNN model have a good effect on image feature extraction, edge information analysis, image enhancement and image segmentation. In this paper, an image enhancement model called AFC-MSPCNN is proposed for image processing. Aiming at the problems of complex structure of pulse coupled neurons, poor adaptive performance and poor processing results caused by repeated attenuation of image signal threshold in experiments. The improved model reduces the computational complexity and enhances the adaptive capability to some extent. We apply the new model AFC-MSPCNN to image enhancement and experimentally verify its good enhancement effect.

Images taken in dim lighting conditions usually suffer from drawbacks such as weak contrast along with unwanted interference. As a solution to this issue, this paper suggests a dim-light image processing strategy rooted in an improved dual-branch adversarial generative model. This method utilizes a generator network that combines a dim-light processing module and an image segmentation module, along with a patch discriminator network, trained using an adversarial loss function. The experiments demonstrate that the suggested approach for improving low-light image quality produces visually more captivating results when applied to real-world datasets. Moreover, it surpasses existing methods in terms of both preserving image structure and reducing noise.

In recent years, driven by the rapid advancements in artificial intelligence technology, image segmentation methods have garnered considerable attention from researchers worldwide. Medical image segmentation, a crucial process, involves accurately extracting structures or regions of interest from medical images. This aids doctors in identifying and quantitatively analyzing various lesions, tissue structures, and organs, thereby playing a pivotal role in medical image analysis and diagnosis. Among these methods, pulse-coupled neural networks (PCNN), leveraging neuronal activity characteristics, have shown significant promise for application. Hence, this paper proposes an optimized FC-MSPCNN (OFC-MSPCNN) image segmentation method, building upon the fire-controlled MSPCNN (FC-MSPCNN), with the aim of simplifying parameter settings. Experimental results demonstrate that this method can notably reduce detection costs while simultaneously enhancing detection accuracy, offering an effective approach for medical image segmentation.

In recent years, there has been a significant amount of research into developing image segmentation methods using neural network models, due to their successful application in the visual field. In this paper, we propose a parameter-continuous fire-controlled MSPCNN (PCFC-MSPCNN) model for image segmentation, which is based on the FCMSPCNN and simplifies related feature parameters. The experimental results indicate that these improvements make the dynamic threshold adjustment of the PCFC-MSPCNN model more appropriate and can provide more accurate results.

In recent years, pulse-coupled neural network (PCNN) has been widely used in the image segmentation field and has obtained relatively satisfactory results. In this paper, based on fire-controlled MSPCNN(FC-MSPCNN), an image segmentation method based on saliency-guided FCMSPCNN(SGFC-MSPCNN) is proposed, which enhances the reasonableness of the variations between the internal activity term and dynamic threshold for the proposed model. and after experimental validation, the method significantly improves the detection accuracy while reducing the detection cost, and thus it is confirmed to be an effective medical image segmentation method.

To address the issue of low segmentation accuracy and high computational complexity in traditional pulse-coupled neural network (PCNN) for medical image processing, this paper proposes an image segmentation approach based on unit-linking fire-controlled multi-scale pulse-coupled neural network (ULFC-MSPCNN). This method simplifies the configuration of data parameters and reduces the number of iterations during the effective transmission period. Experimental results demonstrate that the proposed method significantly reduces detection costs while improving detection accuracy. Moreover, it notably diminishes the randomness and unpredictability of discharged neurons. Thus, it can be seen as one of the effective approaches for medical image segmentation.

Many sectors are challenged by how to effectively represent knowledge in files that contain multiple images closely related to text, and how to make models understand the relationship between images and text. Contrastive Language-Image Pre-training (CLIP) and Bootstrapping Language-Image Pre-training (BLIP) acquire the capability of understanding the image-text relationship through large-scale model pre-training. CLIP not only considers images and their related text but also contrasts images with massive irrelevant text, to improve its capability of generalizing the relationship between images and related text. BLIP enhances its understanding of complex image-text relationships by pre-training and fine-tuning matched image-text pairs. This paper presents an image-text fusion algorithm based on CLIP and BLIP, which gives an accurate and consistent picture of the image-text relevance by fully using CLIP’s image-text generalization capacity and BLIP’s capacity of understanding the complex image-text relationship.

In remote engineering sites, weak or even no signal and inadequate short message communication transmission capabilities fail to meet the requirements for image transmission, leading to the inability to transmit on-site images. The transmission of key images for infrastructure projects in remote mountainous areas without network signal coverage has long been a technical challenge. To address the above issues, this study investigates the image transmission technology based on the BDS-3 RDSS (Short Message Data Service) link. Through the RDSS image transmission protocol, the frame format of the image transmission protocol for BDS-3 satellite RDSS is designed, including communication primitives, communication processes, and interface specifications between upper and lower layers. A data compression and packet retransmission strategy based on the RDSS transmission protocol is proposed, and a package splitting and packaging scheme for BDS-3 short messages is adopted to achieve image transmission with high error correction capability and high compression ratio under narrow bandwidth. Test results show that the proposed technology can achieve image transmission in engineering sites without public network.

In the domain of spine 2D/3D medical image registration, the use of a registration plate is essential for assisting the registration process. To enhance registration accuracy, it is crucial to repair the registration plate images contained within X-ray images. Addressing this challenge, we introduce a GAN-based Mask Inpainting Image Net (MII-Net) model specifically designed for repairing orthopedic locator images in spinal X-ray photographs. This model leverages the Fast Fourier Convolution to reconstruct the image receptive field and employs a strategy that generates a 50% occlusion mask image for training the repair network. The network was trained using custom training and validation sets, and its repair performance was assessed using the FID and LPIPS indices. Post-training results indicated that the model achieved an FID index of 1.45 and an LPIPS index of 0.12 in the validation set, with the SSIM index reaching 0.95. The MII-Net model, powered by GAN technology, guarantees high-quality image repair outcomes, enabling precise repairs of registration plate images in spinal X-ray images. When compared to other models using the same dataset, the MII-Net model surpasses them in terms of both the FID and LPIPS indices.

Medical image segmentation has an important role in medical diagnosis, clinical and other social reality scenes, and it is very important significance to reasonably analyze medical images. pulse-coupled neural network (PCNN) has great potential for artificial neural networks with biological background. In this paper, we propose a medical image segmentation method based on a Global fire-controlled MSPCNN (GFC-MSPCNN) in terms of the influence of neurons. The proposed method improve the interaction between the central neurons and their near-neighboring neurons, which is more suitable for human visual characteristics. It is evident from the experiment that the proposed method can obviously reduce the detection cost and improve the detection accuracy, which is an effective medical image segmentation method.

Deep neural networks have demonstrated remarkable performance in medical image segmentation tasks. Nevertheless, their efficacy often comes at the cost of high model complexity and sluggish inference speeds. To address these limitations, we explore a new channel-level and pixel-level relational knowledge distillation (CPKD) method to efficiently transfer the abundant relational knowledge from the teacher network to the light-weight student network. Concretely, CPKD first captures the inter-class and intra-class semantic information of the teacher network at each channel and pixel, and then distills it to the student network in the form of knowledge. Channel-wise relational distillation utilizes the correlation between channels, while pixel-wise relational distillation utilizes the correlation between pixels. These two forms of relational knowledge complement each other in the learning process, which enables the student network to better simulate the behavior of the teacher network. Extensive experimental results on the medical image segmentation task show that CPKD can significantly improve the performance of the student network compared to other knowledge distillation methods, is highly sensitive to edge information, especially in HD95 metrics our method is far superior to the comparison methods, and has smaller model complexity at the meantime.

Knowledge distillation is an effective model compression and performance enhancement method to instruct students by delivering soft labels or features of the instructor's logits. However, different organ classes in medical images may have similarities in shape, size, texture, etc., which may lead to mutual interference between each class. To solve this problem, we propose Class-wise restructured knowledge distillation (CWRKD). CWRKD generates class-wise features by coarse segmentation prediction of the auxiliary segmentation head with the features of the backbone, and then reconstructs it using our proposed Res module, and utilizes the instructor's features for its to bootstrap it, and also bootstrap the soft labels generated by the auxiliary segmentation head through the teacher's soft labels as a way to reduce the interference problem of similarity between each class. Experiments on different datasets show that CWRKD is effective.

Begin Single-image depth estimation holds significant application value in the field of computer vision. Traditional methods have certain limitations in accurately locating distant regions in images and predicting the precision of object edges. Therefore, this paper proposes a depth estimation method from a single image using fusion of multi-scale features. Firstly, we construct a feature extraction module that combines Transformer and multi-scale feature fusion mechanisms. This module effectively captures features at different scales and levels in images while focusing on global features, resulting in a more comprehensive and precise feature representation. Secondly, we design a Scene Feature Encoding Module that utilizes dilated convolution blocks and adaptive average pooling to extract context correlations between pixels in images, enabling effective recognition of distant regions. Finally, we design a Strip Spatial Perception Module to refine the perception of depth changes, thereby enhancing overall estimation accuracy. Experimental results showcase the outstanding performance of this approach across diverse scenarios, proving its practicality and wide-ranging application prospects in the field of depth estimation from a single image.

With the introduction of blockchain technology and the development of virtual currencies, virtual currencies, which are based on cryptography to ensure the security and anonymity of transactions, have received a lot of attention. As a new payment method that is secure, decentralized and easy to transmit, virtual currencies have also attracted a large number of illegal users and illegal transactions. In order to identify gambling transaction behaviors and gambling-related addresses in the virtual currency market, this paper proposes a gambling transaction feature extraction method based on community detection and network embedding techniques, which obtains a network vector representation of this transaction network structure by discovering a high modularity and highly structured transaction network in gambling address transactions and performing node embedding and averaging calculations based on the node2vec algorithm to complete the extraction of transaction features of gambling addresses and solve the data imbalance problem of the huge gap between the number of historical transactions of different addresses. Finally, based on the feature dataset and several classical machine learning classification algorithms, a binary classification model is trained and evaluated to identify gambling transactions and addresses. Experiments show that all classification models achieve an accuracy rate of 0.72 or higher with high quality of transaction feature data, with the lightGBM model getting the best result of 0.84 accuracy rate, as well as 0.92 and 0.87 recall and F1 scores.

Software-Defined Network (SDN) represents a novel architectural paradigm that effectively separates network control from data forwarding. Due to its global view and programmability advantages, SDN is increasingly being deployed in data centers. However, conventional traffic scheduling methods in data center networks are unable to fulfill current service requirements and fail to effectively utilize the redundant bandwidth within data center network topology, thereby causing network congestion and link load imbalance. To address these challenges, this paper proposes a Bipopulation Fruit Fly Optimized Scheduling Algorithm (BIFOSA). BIFOSA aggregates elephant streams from the same source and employs a modified two-population fruit fly optimization algorithm, thereby ensuring uniform scheduling. Empirical evaluations show that BIFOSA significantly reduces maximum link utilization and improves network load balancing compared to ECMP and Hedera algorithms.

Time synchronization is essential for ensuring the accuracy and timeliness of cooperation and communication between nodes in wireless sensor networks. However, in high-dynamic environments, the high-speed movement of sensor nodes may result in asymmetric wireless transmission links, posing a challenge for time synchronization. To address this problem, we proposes an algorithm that combines a node motion model with time synchronization and validates its effectiveness in improving clock parameters accuracy through experiments. The results show that compared to time synchronization algorithms with symmetric links, this algorithm exhibits superior applicability and synchronization accuracy in high-dynamic environments and can effectively reduce overall communication overhead.

At present, procuratorial organs are facing the problem of diverse and complex types of cases when they initiate public welfare lawsuit. The current research on case classification are not enough to meet requirements of procuratorial organs for future efficiency of public welfare lawsuit. For these problems, we propose case classification model based on BERT-CNN to solve them. First, for the ultra-long sequence input to the model, a dynamic programming algorithm is used for block processing, and then the MemRecall method of CogLTX is used to extract the text blocks that can complete the task and reassemble a new sequence as input. Secondly, BERT, as the word embedding layer, extracts features from the input large-scale source corpus to obtain rich semantic information and converts it into vector output. Finally, the CLS vector output by the BERT model is used as the input of the CNN, which is convolved and pooled, and then spliced into a fully connected layer, and the case is classified through softmax. This thesis fine-tunes the BERT model to make it perform best on case classification. Experiments on the Legal AI Challenge dataset indicate the case classification algorithm based on BERT-CNN is an improvement over others. its F1-score reaches 95.32%, which improved 1.44 percentage points.

Cluster analysis is an important method in data mining and has been widely used in various aspects of real life. The Fuzzy K-Modes algorithm is one of the most commonly used algorithms for clustering categorical data due to its simplicity and effectiveness. However, this algorithm has three main problems: sensitivity of initial cluster centers, vulnerability of distance measurement methods to noise, and failure to consider different contributions of attributes during the clustering process. In this paper, we overview some existing improvement algorithms based on the three problems in detail and conduct experimental analysis of these algorithms on real data sets. Furthermore, we summarize the challenges faced in the process of improving the Fuzzy K-Modes algorithm and give possible research directions in future.

User side energy storage system is generally discrete distributed energy storage system, which can effectively play a role in the power storage in the traditional power systems. The distributed energy storage system can be analyzed theoretically through the classical single capacitor analysis theory, and the complete solution of constant power discharge of single capacitor can be obtained through system derivation. After calculation and simulation, the charging and discharging characteristic curve of group user side energy storage photovoltaic system can be obtained. For photovoltaic characteristic curves, this article collected typical power system characteristic load curves and found that the photovoltaic output curve presents an inverted "U" shaped distribution, but the electricity consumption curve presents an irregular distribution feature. And based on Hop-field Lagrange neural network, a single photovoltaic cell capacitor is used as a neural network node, and the daily distribution curves of photovoltaic cell are obtained through continuous iterative optimization. The optimization result is to optimize the power load curve from the traditional "M" type characteristic curve to a stable horizontal curve. At the same time, the spontaneous self use ratio of photovoltaic power generation was optimized, and the thermal diagram after installing photovoltaic energy showed the full day self use ratio of photovoltaic energy storage. The research results of this article can provide theoretical and practical basis for the economic scheduling of energy storage on the user side. At the same time, for the first time, the Hop-field Lagrange neural network is matched and optimized with the photovoltaic output power generation curve, and the photovoltaic power generation convergence load curve is effectively obtained through hundreds of iterative calculations.

The fluctuation of annual disaster losses is significant, and it is crucial to scientifically evaluate the annual disaster risk in order to improve disaster prevention, reduction, and relief efforts. This paper attempts to establish an annual disaster probability risk assessment method based on historical disaster situations. The main conclusions are as follows: 1) Based on the lognormal distribution, Rayleigh distribution, Weibull distribution, exponential distribution, logarithmic logistic distribution, and information diffusion techniques, a comprehensive annual probability risk assessment method integrating optimal distribution fitting, information diffusion, and risk assessment is established. Using data on annual deaths and collapsed houses caused by disasters from 1990 to 2016 as an example, the feasibility and operability of this method are verified. 2) The extreme values of annual data may have some impact on the assessment results, but the impact of extreme values is more significant. In this study, the assessment results of annual probability risk assessment of death population were not ideal when extreme values were involved, while the impact of maximum values in the annual probability risk assessment of collapsed houses was small. The purpose of this study is to provide case studies and method references for conducting annual disaster risk assessment in related research and business practices.

Facing the problem that the potential complaint risks existing in a large number of service work orders have not been fully explored, this paper puts forward an intelligent analysis method of power supply service demands based on big data. According to the support of the data set, the semantic analysis and PageRank algorithm mining protocol are followed to preprocess the theme information of multi-dimensional user demands; According to the pre-processing results, the step flow of multi-dimensional user appeal theme identification is designed to realize the research on multi-dimensional user appeal theme identification method. The experimental results show that the recognition rate and recall rate of the proposed method are high, which can ensure good and stable recognition results and ensure recognition accuracy and efficiency. Therefore, the multi-dimensional user appeal theme identification method is practical. So PageRank can be applied in multi-dimensional user attraction theme recognition.

The power backbone network can be categorized into a transmission network, a data network, and a support network. Every network is equipped with an extensive array of communication and transmission equipment, produced by diverse manufacturers and consisting of intricate types of equipment. Locating the cause of a fault and identifying the specific equipment that is causing the problem can be challenging in the absence of a relationship map. Therefore, research is currently underway to construct a knowledge graph for correlation of power grid operation faults, utilizing a multi- heterogeneous data fusion method. By correlating knowledge, which encompasses fault cases, fault types, fault causes, and solutions, a logical relationship network is established to tackle issues pertaining to challenging fault cause localization, undefined fault impact scope, inconvenient fault information retrieval, and the inability to extract fault occurrence patterns.

Aiming at the phenomenon that the aircraft fault text data have not been sufficiently exploited, this paper proposed an intelligent aircraft fault diagnosis model based on natural language processing methodology. The proposed model firstly used word2vec to generate 100-dimensional word vectors for every word. Then the extracted text features were input into the LightGBM classification model for further fault diagnosis. Moreover, the Borderline SMOTE algorithm is used to make up the shortage caused by the imbalanced data. We conducted validation experiments on a real-world dataset that recorded by a long-term maintenance and support work. The experimental results indicates that the overall performance of the proposed model is better than other comparison models, which is of significant guidance and reference for the maintainers to take actions to make the aircraft return to normal.

Extreme imbalance in long-tail datasets poses huge challenges for deep neural networks model and visual recognition. Networks models are very easy to be dominated by the "head" class and produce overfitting. At the same time, the ability of the model to model the "tail" class is extremely limited, which leads to serious model performance degradation. Most previous works attempt to address classes imbalance by assigning more weights to the tail classes, and these methods can cause the model to overfit the partial tail classes by directing the network to pay more attention to the tail classes. Although the model performance is improved, the embedding space of tail classes is not really expanded the input features are still biased, the problems of decision boundary shifting to tail classes and the tail class embedding space distortion are not solved. In order to address this problem, this paper proposes a Gaussian perturbation adjusted logit method to improve the decision boundary migration problem, and calibrates the embedding space distortion of tail class by giving different adjusted strength for different classes. At the same time, we study the adjusted strength and the correlation between the model performance, found that giving proper adjusted strength for different classification abilities could be a better solution for the class imbalance problems. The proposed method has been shown to be effective through our extensive experiments on popular long-tail datasets, and it has achievedstate of the art performance on multiple long-tail datasets.

Utilizing the pre-training model and the Text-CNN method, a classification system is constructed in order to propose a multi-label classification solution for terrorist news texts. The ultimate goal is to divide coarse-grained large categories into more detailed subcategories. The data set is selected from the Global Terrorism Database (GTD), and three labels, namely attack target, weapon type, and attack type, are selected in order to classify the text. Introduce a profound deep learning architecture by implementing ERNIE for extracting essential textual characteristics. Subsequently, the architecture is trained with a Convolutional Neural Network, which consists of diverse convolution kernel dimensions. The model suggested in this study surpasses other deep learning models concerning the GTD dataset. It has the potential to refine the granularity of text classification results, simplify subsequent research, and demonstrate substantial practical significance.

To address the issue of insufficient extraction of local key features in existing point cloud semantic segmentation methods, we have made improvements to the RepSurf baseline model. Firstly, in the initial stages of the network architecture, we introduce a multi-layer perceptron to map the input point cloud to a higher dimension, thereby expanding the theoretical receptive field and matching the channel numbers between the encoder and decoder parts. Subsequently, we embed the InvResMLP into the Surface Set Abstraction Channel De-differentiation module to enhance the extraction capability of local features effectively while avoiding issues such as gradient vanishing and overfitting. Experimental results demonstrate the improved RepSurf, when applied in indoor scenarios with the S3DIS dataset as the segmentation target, achieved remarkable results. It achieved mIoU (mean Intersection over Union), mAcc (mean accuracy), and OA (overall accuracy) of 68.1%, 76.0%, and 89.6%, respectively. Compared to the baseline model, our approach exhibited significant improvements of 1.8%, 1.3%, and 0.6% in mIoU, mAcc, and OA, respectively.

Aiming at the problem of low registration efficiency when indoor home robots face complex scenes and low overlap, a PPFECA-Predator indoor laser point cloud registration network that enhances the geometric features of neighborhood points is proposed. First, the point cloud is sent to the residual block composed of core point convolution KPConv to extract features and sent to the geometric encoding module designed in this article to enrich the local geometric feature information between point pairs; secondly, in the overlapping attention module, the graph After the feature fusion of the convolutional network, the attention ECANet module is added to focus on learning the geometric topology information that makes up the graph; finally, the PPF geometric encoding module and the improved graph convolution module are cascaded decoding operations to form a new geometric information enhancement module GEM,enhances the network’s capture of geometric information. The experimental results show that, by analogy with the benchmark network, on the indoor data set 3dmatch, the registration recall rates are increased to 92.2% and 72.14% respectively, effectively handling registration tasks in complex scenarios such as prominent geometric shapes and low overlap.

In order to solve the problems of low efficiency and poor accuracy of traditional bidding scheme selection, an application of big data optimal selection algorithm in project bidding scheme selection is put forward. In this paper, the whole bidding scheme is processed by big data sampling method to speed up the big data processing process. By clustering the sampling results with Single method, the location of the centroid of the natural cluster of big data is determined. Then, the position of the centroid of the natural cluster of big data is adjusted by using the mean updating method to determine its accurate position. On this basis, the bidding scheme set is divided and clustered to realize the classification of bidding schemes. In order to choose the best bidding scheme, we build a projection pursuit model. Through this model, the optimal bidding scheme can be found. The experimental results show that the clustering error of the proposed algorithm is less than 6%, the accuracy is over 94%, and it has obvious advantages in computing speed.

In recent years, artificial intelligence technology has become increasingly closely connected with various fields. However, due to the high requirements of traditional convolutional neural networks for memory and computing resources, it is relatively difficult to deploy them on mobile devices. Therefore, the demand for lightweight neural networks that can be deployed on mobile intelligent terminals is becoming increasingly urgent. This article is inspired by dilated convolution and GhostNetV2, and proposes a new lightweight module-DDC Block. It uses dilated convolution with a larger receptive field structure, combined with depthwise separable convolution, to generate more feature maps through these inexpensive operations, achieving the goal of lightweight. And introduce a decoupled fully connected attention mechanism to ensure high accuracy of the module. We conducted experiments on the CIFAR-10 and CIFAR-100 datasets and compared them with other neural networks. The results showed that this module not only reduced a significant amount of parameter and computational complexity, but also ensured high accuracy.

Recently, two new emerging biometrics technologies, 2D low-resolution palmprint recognition technologies, have received wide attention. Numerous methods have been developed and proposed for palmprint recognition. Among them, one local descriptor, histogram of oriented lines (HOL), has achieved very desirable performance. However, HOL is constructed on a single scale and only exploits the orientation of the minimum response, and thus may lose some important information. In this paper, aiming to further increase the discriminative power of HOL, we propose an enhanced HOL (EHOL) descriptor jointly exploring the information in multiscale and multi-orientation maps for robust palmprint recognition. Furthermore, the EHOL descriptor is only constructed by orientation information and does not use histogram normalization and block strategies to generate the final histograms. Thorough experiments are conducted on two palmprint databases. The experimental results demonstrate that the recognition performance of the proposed EHOL is obviously better than that of state-of-the-art methods.

Two-wheeler riders primarily refer to drivers of bicycles, electric bicycles, electric motorcycles, and motorcycles. Due to the minimal protection afforded to riders on the road, they are highly vulnerable to injuries and fatalities in traffic accidents. Therefore, the identification and supervision of riders on the road are of significant research importance. To address this issue, this paper proposes a novel rider re-identification algorithm based on deep learning, named the two-wheeler rider re-identification algorithm that combines attention mechanism and local feature fusion (AMLFF-Net). Firstly, we divide the network into global and local branches, where the global branch captures coarse-grained features of the entire image, while the local branch captures fine-grained features of specific regions. Secondly, we introduce attention modules and local feature fusion modules. The attention modules enable the network to focus on the target itself, reducing interference from background factors, while the local feature fusion module merges rider features with two-wheeler features, thus increasing the differentiation between feature vectors of different identities. Finally, we compare the proposed method with existing algorithms on two rider re-identification datasets. Experimental results demonstrate that AMLFF-Net achieves Rank-1 accuracies of 72.6% and 89.8% on the BPReid and MoRe datasets respectively, with mAP scores of 73.7% and 91.4% respectively, outperforming other methods and confirming the effectiveness of our proposed approach.

In recent years, the apple growing industry in Qingyang has been seriously threatened by many diseases such as grey mould, rust, brown spot, scar disease and leaf spot. These diseases have caused significant losses to the local economy. This study discusses the method of combining computer technology with deep learning to accurately diagnose these diseases, with the aim of reducing their negative impact on agricultural development. To this end, we collected early disease data sets of apple leaves in Ningxian Modern Agricultural Industrial Park in Qingyang City, Gansu Province, and Haisheng Apple Planting Base in Yulinzi Town, Zhengning County. Considering the problems of low accuracy of classification of apple leaf diseases, difficulty in collecting data sets and huge model parameters, this paper selects grey spot, rust, brown spot, scarring and leaf spot as the research objects, and proposes a MobileViT model suitable for small sample size based on the theory of deep transfer learning. The model aims to solve the problems of large model, low precision and small sample in the process of apple leaf disease detection under complex background. Firstly, MobileViT, Vision Transformer and Swin Transformer are used for the training of the model transfer learning. The experimental results show that the accuracy rate of the MobileViT model is 97.3%, the loss value is 0.169, the model size is 18.9 MB, and the prediction time of a single image is only 2.6 ms. Furthermore, the MobileViT model is optimised by freezing different training strategies, the migration strategy is the most effective, so the average accuracy of the model in apple leaf disease classification reaches 98.54%, and the loss value drops to 0.125. Finally, we developed a WeChat applet to deploy the trained model, and realised the visualisation of apple leaf disease classification. This innovative application not only improves the efficiency and accuracy of disease classification, but also provides new opportunities for the modernisation and intelligence of agricultural technology.

The ability to visualize the semantic connections between relationships and entities is a powerful feature of knowledge graphs. Unfortunately, it is typically challenging to extract the multi-level information of these relationships and entities from the knowledge graph. We have developed a method that can be used to learn such connections using a graph neural network. Our model effectively captures the hierarchical information in the knowledge graph by mapping entities and relationships into a multi-level graph neural network. Specifically, we design multiple graph convolutional layers and attention mechanisms to achieve multi-level embedded representations of entities and relationships.

Accurate segmentation of retinal vessels in fundus color photos holds significant importance in diagnosing ophthalmic conditions. This process plays a crucial role in identifying and analyzing various eye diseases. However, the retinal vascular structure is highly complex, with imbalances in multiscale and foreground-background proportions, making automatic segmentation challenging. Studies have indicated that valuable information is contained in the three RGB channels of color retinal vessel images. Nevertheless, there has been no research to date on integrating the information from the RGB channels with U-shaped segmentation networks for training. Therefore, in this paper, we propose an Attention U-Net with adaptive fusion of RGB three-channel features. The selective channel attention fusion module can adaptively integrate RGB three-channel features, and the fused features are input into the U-shaped network for training. In addition, an Attention Gate module is added before each upsampling concatenation operation in the decoder stage of the U-Net network. This effectively highlights specific target areas while suppressing irrelevant regions of the input image. Experimental results demonstrate that our method outperforms the segmentation results of other algorithms in three datasets in recent years.

Fine-grained image classification has the characteristic of large inter-class differences and small intra-class differences. The challenge lies in the fact that different subclasses have similar structures and minimal differences. The key to distinguishing different subclasses lies in differentiating certain components of the image subject, such as the beak, claws, and tail of a bird, and there is also a certain relationship between these components. With the application of Transformer in visual processing, researchers have discovered the inherent advantages of Transformer in establishing dependencies. This paper fully considers the characteristics of Transformer in mining contextual relationships and proposes a fine-grained image classification algorithm based on Transformer and patch-selection mechanism for extracting component features.

Knowledge distillation requires pre-trained teachers, while self-knowledge distillation can perform knowledge distillation without pre-trained teachers. Therefore, this method is widely used in medical image classification. In this paper, we propose a reverse self-distillation Re-SKD training framework for dermatology image classification. Traditional self-knowledge distillation uses shallow networks as students and deep networks as teachers to guide teachers. The reverse self-knowledge distillation framework proposed in this article uses the shallow network as the teacher and the deep network as the student. Since the shallow layer can capture more shape and edge information in the image, and the deeper layer can learn more semantic information, we utilize features extracted from the shallow layer to construct a teacher to guide the deep teacher. We conducted experiments on the HAM10000 public data set and proved that our proposed training framework Re-SKD can significantly improve the performance of the model in skin disease classification.

Localisation detection and recognition of licence plate plays an important role in road traffic environment. In order to improve the intelligent traffic system and solve the problem of low recognition rate due to tilted licence plate numbers, a licence plate recognition method based on image rectification and improved LPRNet is proposed, and a Chinese parking dataset is used for training and testing. Firstly, the licence plate image is greyscaled, the edges are extracted by Canny operator, the tilt angle is obtained by Radon transform, and the tilted licence plate is cut after tilt correction; then it is recognized based on the model SE-LPRNet. The results show that the average recognition accuracy of the licence plate recognition model is more than 98%, and the proposed method shows strong robustness in the scene of tilted licence plate, and there is also an obvious advantage in recognition speed.

Internet traffic analysis is the core approach to network management and security. In the rapidly changing environment of encrypted traffic, traditional plaintext-based analysis methods have become obsolete. Although there are currently some methods for analysing encrypted traffic, they overlook the inherent logic and hierarchy of different encrypted traffic analysis requirements and lack research into the essential characteristics of encrypted traffic. This article proposes a framework FM-ENet based on Graph Neural Network and Deep learning for Tor encrypted traffic classification to meet the practical needs of network management and security monitoring. A Flow Information Graph has been designed on the basis of Graph Neural Network, which can effectively alleviate the problem of concept drift. In the area of deep neural networks, we have developed an end-to-end multi-level spatio-temporal feature fusion enhanced network module ST-FENet, which has the advantage of automatically learning the non-linear relationship between input and output data without the need for manual feature extraction, this structure can complement each other in classification performance and solve the problem of low classifier efficiency. We compared FM-ENet to current popular methods using the public ISCX-Tor dataset. FM-ENet can achieve higher performance while saving costs, with a 9% improvement in accuracy compared to FlowPrint and a 4.4%, 4.5% and 4.4% improvement in PR, RC and F1 indicators compared to TSCRNN.

Fast and correct identification of water body information from remotely sensed images is important for water resource management and disaster prevention. In this paper, the semantic segmentation network for water body recognition is improved based on the encoder-decoder structure, specifically, based on the DeepLabv3+ model, HRNetV2 model was added to the encoder instead of the original backbone network, and a new attention mechanism CBAM was added to the network model to better obtain channel relationships in the image. This study conducted experiments on water body data using the improved model. After comparing the results with other networks, it was found that the improved model in this study has better accuracy in water body recognition, further improving the effectiveness of identifying water bodies from remote sensing images. It is of great significance for quickly and accurately extracting water body information, flood analysis, and water resource management.

The function of recording and reproducing facts according to statistics is widely used. However, the appearance of image editing software makes it easy to tamper with image information, which reduces its credibility and authenticity. Therefore, judging and restoring the originality, authenticity, and integrity of digital images has become an urgent problem to be solved. To address these issues, we propose a corresponding image recognition technology based on fine registration. We measure the system of image matching and automatic correction for users in different shooting angles, image sizes, image acquisition devices, image rotation differences, and different lighting conditions. This technology aims to eliminate noise pollution in the process of image transmission and establish testing standards for images. Through the deep learning process, we have achieved a high-precision image homology recognition method that possesses the characteristics of good robustness, fast response speed, and strong practicability.

The ability to recognize emotional states, such as happiness, from electroencephalography (EEG) signals has gained significant attention in recent years due to its potential applications in affective computing, mental health monitoring, and human-computer interaction. This paper presents a comprehensive overview of EEG-based recognition of happiness, including signal acquisition, preprocessing techniques, feature extraction methods, classification algorithms, challenges, and real-world applications. Additionally, it discusses the future directions and potential impact of EEG-based happiness recognition.

In order to solve the problem of Qt graphics display performance, this paper proposes a method of accelerating Qt graphics display technology based on domestic GPU Blit block copy. This method overrides the blit block transfer copy function of domestic GPU under the embedded Qt graphics system, and uses Qt's QRegion ability to calculate the list of rectangular vectors of the dirty area to be redrawn in the Qt window system, perform DMA block copy of image source data to target video memory, and calls the domestic GPU function to immediately drive the data content in VRAM for refresh display. The experimental results show that the proposed method can significantly improve the performance of Qt graphics display.

Image defogging methods for traditional physical models produce artifactual effects while removing haze because the coupling of different color channels is not considered. To solve this problem, an alternating least squares post-processing method for variational defogging is proposed. On the basis of the variational model for haze removal, The dehazed image is processed by alternating least squares, which is based on the atmospheric scattering model, and the discrete cosine image transform is used to process the image through transform the image into the frequency domain, and the smoothness of the function is realized by the linear combination of the first few terms of the discrete cosine transform. The transmittance of the three RGB color channels is coupled to effectively remove the artifacts present in the image, and experiments are conducted in the LIVE Image Defogging database to compare with the classical defogging methods. The results of our experiment show that the present algorithm not only reduces the artifacts of the image, but also improves the image quality. It shows that the method in this paper has excellent effect of removing fog.

Recently, Vision Transformers (ViTs) have demonstrated highly competitive performance in image classification and recognition. This is due to the fact that ViT uses a Transformer-based network, which is able to understand the image globally through its long dependency property, and utilizes the attention mechanism to extract stronger features. ViT has proven to perform better than earlier Convolutional Neural Network (CNN) based methods. As a result, despite requiring less computer work, results that are comparable to those of earlier techniques can be obtained. In this work, we explore the possibility of applying this modeling technique to the domain of text-guided image editing. To achieve this goal, we enhance current image generation techniques by utilizing a generator network that is based on the Swin Transformer architecture. By adding a semantic modulation module, this network creates higher resolution images and extracts more feature information from the latent space. We thus introduce the AutoFace technique, which combines the high-resolution image production function of the SwinStyle generator with the image text semantic alignment function of the CLIP pre-training model. This approach enables the generation of higher-resolution images with lower graphics memory and computational performance. The method can generate higher resolution images using less graphics memory and computational power. It can also automatically adjust parameters based on the semantic modulation module for different types of images. Compared to previous methods, the model can be trained in a shorter amount of time. We tested our method on CelebA-HQ and FFHQ datasets. Both the visual and numerical results demonstrate the effectiveness of our method in generating semantically accurate and visually realistic images.

To address the issues of complex backgrounds, imbalanced samples, and inconsistent sample scales in remote sensing image segmentation, this study proposes a remote sensing image segmentation network based on Swin Transformer (SW). The backbone network is designed based on Swin Transformer and incorporates a Mixture of Experts (MoE) structure to separate the model's parameter space. Different scales and scenes of remote sensing images activate different expert models for inference, addressing the complex backgrounds and imbalanced sample issues. By adding a channel attention module to the decoder to aggregate spatial and channel information weights of remote sensing images, the multiscale information of buildings in remote sensing images is effectively utilized, thus reducing the loss of image details during training. Experiments conducted on the public remote sensing semantic segmentation dataset NAIC validate the effectiveness of the proposed algorithm, which achieves mean Intersection over Union (mIoU) and F1 scores of 84.06 and 90.12, respectively, outperforming DeeplabV3+ by 3.44 and 3.77.

In public places, rapid and accurate behavior recognition helps to monitor the occurrence of dangerous events. With the rise of neural networks, many methods have been developed to handle behavior recognition problems. Unlike target recognition, behavior recognition requires not only analyzing the spatial dependencies of the target, but also analyzing the historical information of target changes. We have inherited the traditional dual stream recognition method, in order to better extract spatial and temporal features in consistency, strengthen the contextual information extraction ability and channel fusion. In this paper, we propose a parallel convolutional network (MFC network) for extracting richer contextual information from dual flow branches. In addition, we propose a spatiotemporal modal interaction module (ST block) to improve the robustness of the network. The experimental results on datasets HMDB-51 and UFC-101 show that the method is effective.

The granularity of ore particles is crucial for the efficiency and sustainability of steel smelting operations. This study introduces an innovative segmentation technique for ore particles utilizing a depth camera to capture point cloud data on conveyor belts, which is then processed into a two-dimensional image for analysis. By employing a novel dual-encoding neural network structure, Swin-FUNet, for semantic segmentation, followed by morphological operations and concave point segmentation, the method significantly enhances the accuracy of particle segmentation. Comparative experiments confirm the effectiveness of this approach, offering potential improvements in material utilization, smelting efficiency, and equipment longevity for the steel industry.

A technique for constructing a prototype image library has been proposed, which consists of two aspects: basic support and extended support. Basic support includes detection and quality assessment of individual face images, using HOG features and SVM methods, and training VGG models for evaluation. Extended support allows users to upload three face images with different poses and similarities, using Euler angles and residual networks for pose and similarity checks. Experiments show that with these techniques, the construction cost of the prototype image library can be reduced, and the recognition performance of the system can be improved.

In response to the challenge of effectively identifying artificially generated images from real ones, this paper proposes a deep learning-based approach for authenticating images. The proposed method utilizes a combination of convolutional neural networks (CNN) and generative adversarial networks (GANs) to compare and analyze various indicators of images. Experimental results demonstrate that deep learning algorithms can significantly improve the accuracy and reliability of image authenticity identification. The proposed method has significant implications for protecting intellectual property rights and ensuring public safety. The research contributes to the advancement of computer vision and image processing fields and underscores the need for continued efforts to address the challenges posed by artificial intelligence and image generation technology.

Object detection represents a fundamental task within the realm of computer vision. However, achieving object detection in the dark is still a substantial challenge due to the low contrast of images and the lack of large-scale labeled dark image datasets. One possible solution is to transfer the knowledge of the trained model from a normal illumination source domain to a dark target domain by developing a domain adaptation method. Therefore, we propose a domain adaptive feature distillation network (DAFD-Net) to improve the accuracy of dark object detection. Specifically, we guide the feature extraction of source and target domains by adopting the knowledge distillation training strategy of the teacher-student model and introduce domain adversarial loss and feature distillation loss to jointly conduct semi-supervised optimal learning of the network. Experimental results demonstrate that our DAFD-Net achieves superior performance on the semi- supervised dark detection task.

The detection of anomalies in Key Performance Indicators (KPIs) is fundamental to the operation and maintenance of microservices. The spatio-temporal interaction feature show how time and space relate in a specific context, which reveals KPI data change patterns. However, most existing studies concentrate on the spatial and temporal dependencies of KPI data, neglecting the extraction of spatio-temporal interaction features. In this paper, we propose a Deep spatiotemporal Fusion-based KPI Anomaly Detection method, DFAD, which constructs spatio-temporal interaction features by correlating spatial and temporal features. First, the model utilizes parallel spatio-temporal encoders to capture various temporal and spatial features through convolution-attention. Second, a set of parallel spatio-temporal decoders are designed to independently handle decoding while merging complementary features using cross-attention. Lastly, features are combined from a spatial viewpoint using channel-attention, enhancing the fusion effect. We conducts comparison experiments on four public datasets with various state-of-the-art methods and the results confirm that DFAD enhances detection performance compared to mainstream methods, validating the effectiveness of this approach.

YOLOv5 has been improved to solve the problems such as low speed and accuracy of pedestrian detection task and insufficient feature fusion in complex scenarios. First, the GhostBottleneck module replaces the existing BottleneckCSP submodule for algorithm backbone networks, which improves the detection speed. Then, CBMA module is introduced to make the model enhance the ability of accurate positioning of small targets to improve target detection capability. Finally, SIoU_loss is used as the boundary box regression loss function to effectively predict the distance between the boundary box and the real boundary box, and improve the detection accuracy of the algorithm. The experimental results indicate that the improved algorithm can effectively improve the accuracy of pedestrian target detection.

With the development of deepfake methods, a large number of deepfake images and videos have been widely disseminated on the internet, raising public concerns about the authenticity of information. Therefore, deepfake detection has recently become a hot topic in the field of computer vision, and many methods have been proposed. Currently, frequency-based detection methods have achieved commendable results, but there are still two issues: a) These methods use fixed filters to focus on fixed frequency bands and areas, making them easily distracted by irrelevant information and lacking flexibility for different forgery methods. b) The methods that fuse frequency domain information with RGB information using CNNs do not consider global relationships, so they are insufficient to fully utilize both types of information. To address these issues, we introduce a Frequency-Enhanced Transformer Network (FETNet). Specifically, we propose a Frequency Feature Enhancement Module (FFEM), which is a learnable module capable of flexibly enhancing important frequency bands and regions in the original frequency features. Additionally, we present a Feature Fusion Transformer (FFT) that considers global information to fuse features from the RGB and frequency domains, achieving a more comprehensive feature representation. Through extensive experiments on the FF++ dataset, the effectiveness and superiority of our approach have been demonstrated.

With the development of artificial intelligence technology, the thresh- old of information tampering is reduced, and the quality of counterfeiting is increased. More and more fake information is difficult to distinguish. If such false information is widely disseminated, it will have a particular impact on society. It is essential to note that if face information is tampered with and widely distributed, it could lead to social unrest or damage an individual's reputation, resulting in trust crisis.A face tampering detection method combined with a self-attention mechanism is proposed to solve that traditional residual networks cannot capture the long-range dependencies between video frames and ignore local critical information. The face regions in the video are extracted firstly. Then the spatial features are extracted by the residual network, and the local critical information is learned by the self-attention block. Finally, the long-range dependencies of the spatial features are captured using the gated recurrent unit classification module. Experiments are conducted on the Face-Forensics++. The detection accuracy of the method is improved compared with several baseline algorithms, indicating that the methprocess effectively detect whether the face region in the video tampers.

An improved YOLOv5s model was proposed to address the issue of poor efficiency in steel pipe weld defect detection. First, the network's feature learning and localization capabilities are enhanced by introducing a coordinate attention mechanism. Secondly, compress the model through the lightweight Ghost module to effectively get fewer network parameters. Thirdly, standard convolution and depth-separable convolution are combined in the Neck module for convolution operations, aiming to reduce model complexity while maintaining accuracy. Finally, introduced a new WIoU loss function to speed up model convergence. Based on a self-created dataset, the detection of three types of surface defects, including overlap, spatter, and undercut, is conducted. The experimental findings demonstrate that the improved model achieves a 2.4% increase in mAP. Furthermore, the model also achieved a 22.4% reduction in parameter count. These findings indicate that the improved model can effectively meet the practical detection requirements. The experimental comparison confirms the feasibility and effectiveness of the proposed improved algorithm for detecting surface defects in steel pipe welds.

In order to solve the puzzles such as missed detection, poor real-time performance, low accuracy, and limited front-end hardware in PCB board surface defect detection, an improved lightweight model based on YOLOv5s is proposed. First, the C3Ghost module is used to replace the C3 module in the model backbone network, secondly the lightweight GhostConv convolution compression model is introduced, Once again ECANet is added to the backbone network to strengthen the ability to extract key information, and finally the image data set is re-clustered with the help of K-means clustering + genetic algorithm. Experimental results show the Parameters and Size of the model proposed in this article are reduced by 32.3% and 33.8% respectively, reducing the dependence on the detection front-end hardware conditions and improving the detection performance mAP@0.5 of the model. Its detection accuracy is high, the research shows that using the improved model to identify and classify PCB surface defect detection has better real-time performance and higher detection accuracy.

Pedestrian detection is one of the most important branches in the field of computer vision. It has great development potential and application prospects. In recent years, the application of intelligent monitoring, unmanned driving and intelligent robots has rapidly promoted the research and development of pedestrian detection. In the dense crowd, the body of a small number of pedestrians cannot be fully displayed, which makes the detection of pedestrians missing. In the reidentification research of specific pedestrian, it is difficult to extract effective pedestrian features due to occlusion, posture and shooting angle. Therefore, in view of the above problems, this paper studies pedestrian detection algorithm to achieve fast and accurate pedestrian positioning, the main research is as follows. In order to improve the accuracy of target recognition of specific pedestrians, based on pedestrian detection, a pedestrian re-identification method based on improved AlignedReID network is proposed. The method combines the DenseNet121 network with less parameters and good generalization performance, and integrates the global and local features of pedestrians to achieve accurate identification of specific pedestrian targets. The experimental results show that on the pedestrian unobstructed Market1501 data set, the Rank- 1 and Rank-5 recognition rates were 93.8% and 97.3%, respectively, and the mAP was 90.5%, which are 1.8%, 0.9% and 2.0% higher than that of the original AlignedReID network.

Graph anomaly detection plays a key role in many real-world scenarios such as social network flooding detection and financial fraud detection. Graph anomaly detection methods based on contrastive self-supervision have been proven to be effective, but current models lack full utilization of features and lack attention to the over-smoothing problem during training. Therefore, this paper proposes a new contrastive self-supervised model for sampled node and subgraph instance pairs, which fully captures the information of nodes and subgraphs, and at the same time reconstructs the information of subgraphs, which effectively solves the problems of insufficient feature capture and over-smoothing of graph neural network modules. Experiments on three public datasets and three benchmark models demonstrate the superiority of our proposed model.

With the development of science and technology, table tennis robots are becoming more and more common in the daily competition and training of athletes. This study is mainly for the research and development of the vision of table tennis robots. This study uses long and short time memory network (LSTM) to predict table tennis trajectory. We collected a large amount of table tennis sport data and used it to train and test our LSTM model. By analyzing the motion trajectory of the ball, we designed an effective LSTM architecture that can capture the complex dynamic properties of table tennis movements. Experimental results show that our model achieves remarkable success in table tennis trajectory prediction, with higher accuracy and robustness compared to traditional methods.

At present, the main method for detecting foreign objects inside high-voltage towers is manual methods. This method is labor-intensive. To address this issue, this paper proposes a detection method for bird nests in high-voltage towers based on unsupervised depth estimation and 3D reconstruction, which can efficiently and accurately detect bird nests on high-voltage towers and obtain more information. The results showed that after obtaining the depth map information of the bird's nest, the three-dimensional reconstruction method was applied to determine the coordinates of the point cloud. Based on these point cloud information, the 3D model of the bird's nest can be restored, which can minimize subjective judgment errors applied in daily inspections. This method provides a new direction for power system detection.

Aiming at the problem that it is not easy to detect motorcycle driving violations on the road, a motorcycle violation detection method with improved YOLOv8 and DeepSORT is proposed. The method uses the YOLOv8 algorithm as a target detection model, introduces a global attention mechanism (GAM Attention) in the backbone network to refine the learning of key semantic information and improve the feature extraction capability of the network; adopts the WIOU (Wise-IoU) loss function to accelerate the convergence speed and improve the regression accuracy; and uses DeepSORT as a target tracking model. ShuffleNetv2, a lightweight network, is introduced as the appearance feature extraction network to reduce the parameters of the model and maintain good accuracy; re-training on the motorcycle reidentification dataset improves the identity switching problem of the motorcycle target in the occlusion scenario. Finally, the tracked motorcycle targets are detected for driving violations. The experimental results show that compared with YOLOv8, the motorcycle and helmet detection accuracies are improved, reaching 98.2% and 98.6%, respectively; the size of the DeepSORT tracking model is reduced by 82%, which enhances the model's portability; and the algorithm has strong robustness in different traffic scenarios. Finally, the motorcycle target's driver and passenger did not wear a safety helmet, adding a parasol, and overloading three driving violations were detected, and the recognition rate reached 98.8%, 100%, and 98.5%, respectively, indicating that the method has some practical value.

Coal mining in open-pit mines can lead to spontaneous combustion due to the accumulation of coal. During the early stages of spontaneous combustion, smoke is generated, and its diffusion characteristics are similar to the dust stirred up by vehicles in the coal yards. The diffusion of smoke not only affects the safety of coal production but also poses health risks to the staff. Currently, most smoke detection algorithms focus on fire backgrounds, including both flames and smoke targets. However, there is limited research on smoke detection in the environment of open coal yards. The unique characteristics of the coal yard environment, such as the easy diffusion of smoke and dust with no distinct boundaries, make smoke detection challenging. The conventional smoke detection algorithms are not suitable for detecting smoke targets in coalfield environments. To address this issue, this paper proposes the M-ST-YOLOv5 model (Mobile-SwinTransformer-YOLOv5). Firstly, incorporates the concept of the MobileNetV2 network architecture to reduce the complexity of practical deployment. Secondly, the Swin-Transformer module is added to enhance the network's recognition capabilities for smoke and dust. Experimental results demonstrate that the proposed method effectively reduces the number of network parameters while improving the smoke detection rate.

In the field of deep learning-based object detection algorithms, the YOLO algorithm has attracted a lot of attention due to its advantages in speed and accuracy. However, there are problems such as low detection speed and poor detection accuracy in the application of UAV images. For this reason, an improved small object detection algorithm MTF-YOLOv5 based on the YOLOv5 algorithm is proposed. Firstly, MobileNet V3 is used to replace the main network to reduce the amount of network parameters and improve the running speed. Secondly, a small object detection layer is added to the original network structure to enhance the detection ability for small objects. Finally, the FocalModulation is introduced to replace the SPP module to improve accuracy. The improved algorithm is tested on the processed TinyPerson dataset. Experimental results show that the mean average precision (mAP) on the YOLOv5s algorithm improved by 4.1%, and the FPS increased from 86 to 136, improving the speed by 51%. While maintaining detection speed, the updated algorithm successfully raises detection accuracy.

This paper proposes an optimized algorithm for tracking maneuvering weak radar targets in complex sea environments. In the proposal, we first preprocesses radar echo images using the fractional Fourier transform (FrFT) to reduce clutter interference. Then, Feedback factor and residual resampling are introduced to optimize the interactive multiple model particle filtering (IMMPF) algorithm to achieve effective tracking of maneuvering targets. Additionally, the optimized IMMPF is combined with track before detect algorithm (TBD) to suppress clutter and accumulate target trajectory. Simulation experiment results demonstrate that the proposed algorithm can effectively track targets compared to other TBD algorithms.

In low-illumination environments, the contrast between targets and the background sharply decreases, and imaging sensors introduce complex random noise while capturing more light, leading to severe distortion in the target features of visible light images. Under these conditions, target feature-based detection algorithms face significant challenges. In this paper, we propose a coarse-to-fine change detection algorithm that detects targets in low-illumination environments by focusing on the change features generated by the targets. First, a binary classifier based on ResNet18 is trained with a specially curated dataset of targets and noise. Second, this classifier is applied to identify potential change objects in test data, preserve targets, and extract corresponding bi-temporal local regions of interest. Third, novel difference feature extraction operators are employed to generate local difference images. Next, a Laplacian-of-Gaussian-based graph cut algorithm is used to perform binary segmentation, distinguishing foreground from background in the images. We validated the feasibility of this algorithm in three challenging night scenes. Compared with the current state-of-the-art unsupervised change detection algorithms, the proposed algorithm shows overall better detection performance.

With the advancement of society, the phenomenon of population aging has evolved into a shared concern among nations worldwide. In response to this pressing challenge, this paper proposes the development of a predictive model for identifying risk events among the elderly. By scrutinizing behavioral patterns and instituting early warning mechanisms for potential hazards, the model aims to discern intricate and individual behaviors, thus bolstering the safety of elderly individuals, particularly those residing alone. Leveraging Kinect camera technology, the model captures human skeletal data alongside YuNet and MnasNet for facial information extraction. This amalgamation of data interprets the language of human body status, subsequently translated into natural language through the employed language model, facilitating high-precision real-time monitoring.

There is a problem that it is always difficult for picking and spraying robots to identify the objects for such factors such as leaf occlusion, similar color and others. In this paper, based on the citrus image data set collected from the actual orchard scene, image features are extracted and a mathematical model is constructed to identify and count the number of citruses in each image, and the distribution of apples in the whole data set is displayed. A model is constructed to be used to identify and count citrus via these images with diversity of citrus colors and the complexity of the background. Firstly, the yellow mature citruses are focused on using the HSV (Hue, Saturation, Value) method for the data sets. Secondly, the empirical HSV range for two citrus fruits, cyan and yellow, was defined, and contour processing is introduced to optimize the recognition model to process more citrus image datasets. Finally, such three types of objects as immature citrus, leaves with similar colors, and citrus covered by leaves are distinguish. The research results indicate that we adopt more detailed color threshold settings and morphological operations with the eccentricity of the contour is less than0.85, and can distinguish citrus and leaves through image segmentation, and the recognition effect on citrus objects is improved.

A long-term soil erosion monitoring method based on remote sensing data from the Land Satellite 8 OLI satellite has been proposed. According to the needs of soil erosion calculation, the Chinese soil erosion engineering model is used to calculate the soil erosion modulus. Perform factor calculation analysis to determine factors and calculate annual rainfall erosivity; Obtain soil erosion data and determine the spatial distribution characteristics of soil erosion intensity in the region. By using vegetation optical thickness feature analysis and multispectral remote sensing feature monitoring methods, soil moisture inversion analysis was carried out to extract moisture information parameters, and long-term monitoring of soil erosion based on remote sensing data from the Land Satellite 8 OLI satellite was achieved. The experimental results indicate that this method can improve the accuracy of long-term monitoring of soil erosion.

In this paper, we propose a multimodal pedestrian detection algorithm based on heuristic optimization and deep learning model, which is designed to cope with the challenges of multi-scale, partial occlusion, and environmental interference. The algorithm combines the advantages of traditional feature extraction methods and deep neural networks to improve the accuracy of pedestrian detection. In the heuristic and deep learning multimodal model, first, a cross-scale feature fusion module is designed to improve the detection performance of multiscale and partially occluded objects, including classical features such as appearance and pose as well as some deep features in vector format. The module fuses global features from different residual layers and multi-scale local area features, which improves the multi-scale feature fusion capability of the model and expands the feature acceptance domain of the backbone network. Then, Euclidean distance (ED) and location constrained linear coding (LLC) are introduced to complete the matching problem of the target, decoupling the input features from the channel dimension, which can be used for classification and localization. Finally, according to the error of feature matching, the IOU filtering module (IFM) is applied to refine the target state and filter out some invalid candidate targets, respectively,to learn the real regression parameters, thus simplifying the network structure and improving the model generalization ability. The experimental results show that compared with other pedestrian detection algorithms, the proposed algorithm is more accurate in detecting pedestrian targets in complex environments, and the Ran- 1 is improved by 9.2% compared with the single-stage algorithm.

Real-time monitoring of dense crowds can provide data support for early risk warning, evacuation guidance, and emergency resource scheduling. The YOLOv7 algorithm for dense crowd detection has been improved to effectively obtain crowd information. The YOLOv7 framework has been pruned to create a lightweight network, and CARAFE has replaced the upsampling module for adaptive content awareness to improve small target attention. NWD is used instead of C-IoU to enhance model generalization and optimize loss function convergence. Compared with the original YOLOv7 algorithm, parameter reduction is 50%, AP@0.5 increased by 7.9%, and FPS increased by 9.2%. Results demonstrate that this improved algorithm can efficiently improve accuracy in detecting dynamic videos of dense crowds; results from pedestrian street detection show feasibility and effectiveness.

In the field of Unmanned Aerial Vehicle (UAV) patrol image detection, the challenge of accurately and swiftly detecting images is a significant issue. However, current detection models often fail to achieve a balance between precision and detection speed. To address this problem, we propose a detection model named Uni-YOLOv5s. Our model introduces a module called UniReplkNet into the YOLOv5s model, which utilizes large convolutions to extract image feature information. We evaluated our model on the InsPlAD dataset. The results demonstrate that our model can rapidly detect UAV patrol images and achieve superior performance metrics. The recall rate is 92.7%, which is 2.5% higher than the baseline YOLOv5s, Thus verifying that our method is superior to existing methods.

In fact, it is crucial to accurately identify and distinguish steel surface defects in the metal industry. Aiming at the problems of misdetection and poor detection accuracy in scratch and crack defect detection, this paper proposes an algorithm for detecting surface defects on steel based on the LSwin Transformer fusion model. The algorithm takes Mask R-CNN as the baseline framework, fuses it with the improved Swin Transformer backbone network, combines the feature extraction capability of convolutional neural network with the global dependency construction capability of Swin Transformer, and finally introduces the multi-head local window offset (LS-Tr) module to detect steel surface defects by means of a lightweight PAN and Mask Switch module (MSM) to detect steel surface defects, thus reducing the number of model parameters and increasing the inference speed. Experimental validation on the NEU-DET dataset shows that the accuracy of the algorithm in detecting scratched (SC) and cracked (Cr) defects is improved by 29.8% and 6.9%, respectively, and the accuracy in detecting the mAP value reaches 82.5%, which is 7.1% higher than that before the improvement. It is better than the current mainstream steel surface defect detection algorithms and meets the needs of steel surface defect detection tasks.

Infrared imaging technology has become indispensable across numerous fields due to its unique capabilities. However, small object detection from infrared images is still a challenging task because of the inherent noise and low contrast often present in infrared imagery, and the small target size further increases the detection difficulty. To relieve these problems, we propose a small object detection method based on YOLOv8 with an attention mechanism and multi-level feature fusion from infrared images. Firstly, an attention mechanism is introduced to extract long-range image features, which is effective in decreasing the bad effect of complex image backgrounds. Secondly, multi-level feature fusion is used in the detection neck to recover image details for small objects. Experimental results show that the proposed method is beneficial in improving the detection performance of small objects from infrared images.

Document images often contain critical textual information, and once manipulated or edited, they can cause significant harm to society when propagated on the internet. To date, although some works have achieved research results in the field of document image forgery detection, they have not fully explored edge features, which are important feature information in document images. In response to this research status, we propose a new document image forgery detection method based on multiscale edge similarity (MES). The proposed method employs multiscale difference of Gaussian (MDOG) to calculate the edge structure similarity, edge contrast similarity and edge information entropy at each pixel between the image to be detected and the reference image at different scales. These computed results are combined to generate an edge similarity distribution map, which is used to pinpoint the specific location of the forged area. Extensive experimental results demonstrate that the proposed method outperforms some of the current state-of-the-art forgery detection methods and shows robustness against several common postprocessing operations.

The primary approach to analyzing the model of a maneuvering target is to treat the target's maneuvers as an increase in the state noise variance. Within this context, the Singer model and the "current" statistical model are extensively employed, while the Jerk model is recommended for targets with high agility. This study conducts simulation tests on three frequently utilized models for maneuvering targets. It finds that the Singer model is apt for describing targets moving at a steady speed or with uniform acceleration, whereas the "current" statistical model is particularly effective for modeling target movements involving step changes in acceleration, especially in comparison to the Jerk model. The outcomes of these simulation tests reveal that the proposed method achieves a root mean square error (RMSE) of 3.92 when transitioning from uniform acceleration to step-acceleration maneuvers. This RMSE is significantly lower than those of 6.23 and 4.78, indicating enhanced tracking precision.

In order to solve the problem that traditional wafer localization methods are easily affected by light, noise, etc., and that deep learning target detection methods are slow for wafer detection, we constructed the wafer fast detection algorithm YOLOv5s-fast. firstly, in the feature extraction part, we used the FasterNet network for feature extraction, which accelerated the algorithm running speed; secondly, in order to improve the algorithm detection performance, we added BRA bi-directional routing attention (Bilevel Routing Attention) to feature fusion part of the network to strengthen the feature fusion capability. Secondly, in order to improve the detection performance of the algorithm, BRA Bilevel Routing Attention is added in the feature fusion part of the network to strengthen the feature fusion capability. Finally, the MPDIOU loss function is used to improve the model detection performance. The average detection accuracy of the algorithm is 99.50%, the number of parameters is 1064176, and the detection speed is 2.8GFLOPs, which improves both the detection speed and the detection accuracy, and at the same time realizes the algorithm lightweight.

Traditional highway tunnel management systems often face problems such as independent and decentralized management, inflexible monitoring, and low visualization of disease details. Especially in the face of traffic accidents, lane foreign objects, vehicle congestion, and vehicle and pedestrian congestion in tunnels, traditional warning systems often fail to provide timely and accurate warning information, posing great risks to the safe operation of tunnels. In response to the above issues, this article proposes an intelligent early warning system for highway tunnel safety monitoring data based on Markov model (MM). This system fully utilizes the advantages of MM in describing random processes, and achieves dynamic monitoring and early warning of tunnel safety status through real-time analysis of tunnel monitoring data, combined with Internet of Things (IoT) technology. The experimental results show that the intelligent early warning system for highway tunnel safety monitoring data based on MM has high accuracy and reliability. It can not only achieve real-time monitoring and early warning of tunnel safety status, but also provide rich visual information to help managers better understand the operation status of tunnels and improve the level of tunnel safety management.

In special circumstances, people in power aerial work scenes often have to face high-risk and complex working conditions. How to ensure the safety of workers and improve their working efficiency has become one of the urgent problems to be solved. Therefore, an active tracking algorithm for people in power aerial work scenes based on illumination weight distribution and attention is proposed. The dangerous source structure of power aerial work scene is constructed, and the light weight distribution and attention are combined, and the active tracking index is constructed to evaluate the risk level, which is divided into five levels, and the active tracking offset of personnel in power aerial work scene is calculated to locate the abnormal point of personnel in power aerial work scene. The experimental results show that the response curve of the algorithm in this paper is in good agreement with the ideal curve, and the active tracking accuracy is high, so it has good active tracking effect.

Banks often need to collect and store customers' personal information due to the nature of their business. While this business model provides convenience to customers, it also increases the risk of personal information leakage. The widespread use of monitoring devices has led to rapid development in intelligent supervision technology. This research focuses on the actual work scenario at bank counters, aiming to intelligently recognize tellers' behavior in using mobile phones. Considering the requirements for both model accuracy and real-time performance, we propose our model based on YOLOv8 for the mobile phone detection task. To address the accuracy issues in the original model, the paper enhances the model's learning on hard samples and introduces an attention mechanism to improve accuracy. To tackle the real-time performance issues in original model, the paper optimizes convolution structures and adopts model channel pruning to enhance inference speed. Our model achieves a great balance between accuracy and real-time performance through comparative experiments before and after improvement.

Given the rapid advancement of deep learning in recent years, pedestrian trajectory tracking technology has emerged. Target tracking is used to continuously track detection objects in video sequences to obtain information such as motion trajectories and positions. For pedestrian trajectory tracking, this article introduces multi-target tracking methods based on convolutional neural networks and Transformer, summarizes the improvement points of various algorithms of "one-step method", "two-step method" and Transformer, and organizes the largest pedestrian tracking data set MOT Challenge, evaluation indicators and experimental results point out existing problems in trajectory tracking and look forward to the development direction.

To achieve accurate and rapid fire detection, this paper improves the network structure of YOLOv4 through various means. Given the numerous parameters of the original YOLOv4 network model, we introduce the lightweight MobileNetv3 backbone network to replace the original backbone, thereby reducing network complexity. Additionally, we enhance detection accuracy by incorporating the Inverted-bneck-shortcut structure in the prediction network to facilitate multi-scale feature fusion and prediction. Since publicly available fire image datasets are insufficient for training deep neural networks, we collect fire images through various methods and further expand the dataset using data augmentation techniques, ultimately creating our own fire image dataset. Experimental results demonstrate that the improved YOLOv4 fire detection model presented in this paper is approximately 1/3 the size of the original YOLOv4 model, with a nearly 40% increase in inference speed and a 6% improvement in algorithm accuracy. This suggests that the proposed MobileNetv3-YOLOv4-based fire detection algorithm can accurately and rapidly identify and locate fire targets in images.

Transformer is the key equipment in the construction of the power system, and its safe and stable operation is the foundation of the power system. Aiming at the problems of difficult identification and low detection accuracy of transformer windings in the running state, a fault diagnosis method of transformer windings based on digital twin technology is proposed. Digital twin, is to make full use of the physical model, sensor updates, operation history and other data, integrate multi-disciplinary, multi-physical quantity, multi-scale, multi-probability simulation process, and complete the mapping in the virtual space, so as to reflect the corresponding physical equipment of the whole life cycle process. A high simulation digital twin transformer model based on the transformer entity is established, and then the information changes of the tank surface vibration information of the digital twin transformer under different working conditions and various winding faults are deduced through multiple physical field simulations. The paper studies and analyses the fault diagnosis of transformers based on the support of digital twin technology and focuses on the prediction of transformer faults. It promotes the prediction of transformer faults as well as research based on digital twin technology.

The surface identification of integrated circuit (IC) has not been standardized yet, and the task of detecting IC logos is often carried out manually, which is prone to high subjectivity and low efficiency. Traditional machine learning methods still rely on prior knowledge and handcrafted features, which yield unsatisfactory performance. In order to improve the efficiency of logo detection tasks, this paper constructs a dataset for IC appearance and proposes a deep learning algorithm based on the latest single-stage object detection neural network YOLOv8. By introducing Global Attention Mechanism(GAM) into YOLOv8, the information diffusion is effectively reduced, and the deep information is extracted, allowing the neural network to learn more representative features and further improve the average precision.

To make ensure the quality of service of wireless Mesh networks and improve network performance, this paper aims at the problem of load imbalance caused by several existing main wireless Mesh network load balancing algorithms without fully considering the backbone network nodes and link states. A dynamic load balancing algorithm based on network calculus. First, use the depth-first traversal algorithm to construct available paths; then, establish the queuing delay and data backlog model of nodes, and combine the comprehensive calculation of path reliability with nodes and links to obtain a set of reliable paths; finally, calculate the allocation data of different paths The weight of the packet, and distribute data packets on demand according to the weight. Simulation results show that compared with traditional AODV protocol, HWMP protocol and improved LB-HWMP protocol, the algorithm proposed in this paper is superior in delay and throughput.

The circulation of government data resources is an important link to promote the construction of new smart cities and the modernization of urban governance capabilities. At present, academia and industry use government information resource cataloging system, logical centralization and physical dispersion, and data to the cloud to solve the problem of non-sharing of government data to a certain extent. However, the problem of difficult cross-department and cross-level business collaboration still exists. In order to solve the above-mentioned supervision problems, focusing on the research on the supervision mechanism of rights and responsibilities in the process of cross-departmental circulation of government affairs data, this paper proposes a government affairs data rights confirmation mechanism based on blockchain technology, which is an incomplete decentralized method for central supervision. , and a strong supervision application model based on this method to realize the supervision of data quality and sharing authority. The experimental results show that the proposed theoretical algorithm model has scientific significance for solving the unclear rights and responsibilities in government data sharing, and the supervision of dynamic update of authority. and application value.

With the deep integration of new technology concepts such as big data, cloud storage, the Internet of Things with enterprise management, power grid infrastructure has also entered a period of rapid development, and the realization of safe and reliable data sharing has emerged as an urgent issue that requires resolution. Firstly, the existing attribute-based encryption schemes are mostly deployed in the cloud environment, and it could be challenging for users to achieve fine-grained access control to ciphertext while taking into account the search query function. Secondly, current attribute-based encryption schemes face the single-point bottleneck problem issue brought on by a single authority managing the whole attribute set, making it tough to guarantee data confidentiality and security. In view of the above situation, this paper designs a blockchain-assisted searchable attribute-based encryption scheme, which uses the consortium blockchain's consensus nodes to generate system parameters and user's key through (t, n) secret sharing, besides that, the user's access requests and operations are recorded on the blockchain, realizing the traceability of transactions. The system entirely conceals the access policy and ciphertext's keyword characteristics while also supporting fine-grained access control to users. The security analysis and performance evaluation on the proposed scheme demonstrate its reliability and system-level robustness.

In order to better protect sensitive data within organizations and enhance employee productivity, this study conducts analysis on recent internet browsing behavior data. Different internet browsing behaviors such as browsing duration, browsing time slots, and IP addresses are used as the dataset. A novel encoding format is proposed to establish a one-to-one correspondence between binary digits and Chinese characters. The processed data is then fed into the proposed immune deep neural network for training, aiming to develop a highly accurate model for predicting internet browsing behavior. Through experimental comparisons, the proposed algorithm demonstrates superior prediction accuracy, providing valuable insights for the scientific management of user internet behavior and offering strong support for safeguarding network data security.

When controlling air conditioning, due to the lack of comprehensive analysis of the relationship between energy consumption of different units, the energy consumption loudness of air conditioning is high. Therefore, a study on energy-saving adaptive control algorithm for high heat density computer room air conditioning is proposed. Based on the operational characteristics of air conditioning in high heat density computer rooms, with the objective function of minimizing unit operating power consumption to meet the cooling load demand of the front-end, a high heat density computer room air conditioning energy consumption model was constructed from three perspectives: chillers, chilled water pumps, and cooling water pump pipes. In the control algorithm design stage, the Double Pools based DQN algorithm (DPs DQN) was introduced. Based on the condition of whether optimization of system cooling load allocation is needed, the sample data of high heat density computer room air conditioning operation was divided into two categories: samples that require optimization of system cooling load allocation and samples that do not require optimization of system cooling load allocation, The energy consumption model of the high heat density computer room air conditioning constructed in section 1.1 is used to make action decisions throughout the entire action space, and the result of gradient descent update is the control parameter. In the test results, the designed energy-saving adaptive control algorithm for high heat density computer room air conditioning can effectively reduce energy consumption during the operation phase of the air conditioning system, while achieving stable control of the computer room temperature.

Time Sensitive Software Defined Networking (TSSDN) is a new network architecture that combines Time Sensitive Networking TSN and Software Defined Networking SDN with the common advantages of both. In this paper, TSSDN is compared with different airborne networks from different perspectives and time-triggered traffic delay analysis is performed. Finally, it analyzes the advantages of TSSDN in civil aviation in general and the perspective to be improved.

The purpose of this paper is to propose a quality evaluation model of college students' innovation and entrepreneurship education based on grey relational degree algorithm, which aims to help higher education institutions and educational decision-makers accurately evaluate the quality of college students' innovation and entrepreneurship education. Firstly, through literature review, this paper analyzes the existing problems in the evaluation of innovation and entrepreneurship education. Then, the principle and basic steps of grey relational degree algorithm are introduced. Then, a quality assessment model of college students' innovation and entrepreneurship education based on grey relational degree algorithm is constructed and verified by case data. Finally, the model is evaluated and analyzed, and the future research direction is prospected.

The interesting man-machine game tests people's intelligence and the computing power of machines, making the ancient game of chess more dynamic and becoming the key field of modern artificial intelligence research. In this paper, the minimum value maximum algorithm is used to realize the search algorithm of man-machine game. In the process of human and computer chess, the computer will choose the move method that makes its own situation score the highest, and the human will choose the move method that makes the computer situation score the lowest, which is the basis of the design of the minimum value maximum algorithm. The min Max algorithm also depends on how well the evaluation function is designed. On the basis of the minimum maximum algorithm, the pruning algorithm is further optimized, which can reduce the size of the search tree, optimize the search efficiency, and reduce the chess move time of the computer. After the chess game program is completed, man-machine chess can be realized. The human moves first, and the computer chooses the position of the drop according to the data returned by the algorithm and then chooses the most suitable move to achieve the effect of man-machine chess.

The operation of crane operation commanding personnel not only involves their own safety but also has intricate connections with the lives and property of others. This determines that crane operation commanding personnel need to have higher qualifications and more professional skills, and can only take up positions after passing the examination. However, considering the actual situation, the practical examination of crane operation commanding personnel is still primarily conducted in the form of oral defense, which makes it difficult to achieve objectivity, accuracy, and automation. Based on artificial intelligence, this paper delves into the intelligent scoring system for the practical examination of crane operation commanding personnel, aiming to thoroughly address the issue of automatic judgment of examination and evaluation indicators in the actual operational assessment of crane operation commanding personnel. We collected practical operation data of crane operation commanding personnel and carried out data preprocessing and feature extraction. Using machine learning algorithms, we built a scoring model. Through the integration of cameras, gesture parsing, and learning models, we realized the automatic recognition and evaluation of command gestures of examinees during the examination process.

The economic operation mode of the power grid refers to the operation mode that achieves the minimum power grid loss and maximum benefits under the premise of ensuring the safe, reliable, and efficient operation of the power grid. This article proposes a fully automatic intelligent analysis system for the economic operation mode of the power grid based on RPA+LLM. The system uses Robotic Process Automation (RPA) technology to automatically capture and integrate power grid operation data, and uses Large Language Models (LLM) technology to automatically generate and evaluate analysis reports on the economic operation mode of the power grid. The system has the characteristics of fast data acquisition, accurate data processing, and automatic generation of analysis reports, providing effective support and reference for the optimization of the economic operation mode of the power grid.

To detect corrugated paper quickly and accurately, a YOLOv5-based algorithm has been created. In order to improve the fusion of lower layer feature information, the operation of GSConv layer is introduced in the feature enhancement network algorithm. for better utilizing feature semantic information, the CARAFE upsampling operator is used instead of the original upsampling. MobileNetv3 was used to replace its backbone network to reduce computing costs and achieve lightweighting. Using the corrugated paper dataset constructed by the authors as an example, the improved algorithm was found to have a parameter size of 1.3 M based on the results., computational complexity of 1.8 GFLOPS, and weight size of 2.9 MB. Compared to the original YOLOv5n algorithm, these values decreased by 27.8%, 56.1%, and 21.6%. Its key technical indicators mAP@0.5 and FPS can be as high as 88.0% and 68.5 respectively. These results show that the improved algorithm satisfies the requirements for on-site deployment on mobile devices. The research shows that applying improved algorithms to conduct real-time and accurate inventory of corrugated packaging is a feasible approach.

In recent years, short video platforms have gained prominence as a primary medium for communication. This paper delves into their application within the electric power industry. User-generated labels often fall short of effectively matching video content, providing limited guidance to both content reviewers and viewers. Moreover, manual content review heavily relies on human resources, necessitating a more intelligent approach to label generation. To address this, the paper introduces a Content-Aware Dynamic Graph Convolutional Network, harnessing artificial intelligence for precise multilabel generation in short videos. It begins by implementing an inter-frame absolute difference method with frame intervals to obtain keyframes. Subsequently, a content-aware dynamic graph convolutional network processes these frames, generating multiple labels rooted in content awareness. Ultimately, a set of key information labels is selected according to specific rules. This method significantly enhances label recognition accuracy, surpassing traditional approaches. Experimental results confirm its efficacy in multi-label recognition for short videos, promising advancements for the field.

The dynamic performance (stroke, resolution, speed, etc.) of the focusing mechanism of a microscopic imaging system is the key factor affecting the image quality and detection efficiency. The focusing speed and resolution of the microscopic imaging system depend on the ultra-precise motion of the focusing mechanism. In this study, magnetic shape memory alloy(MSMA) is used to develop a driver for rapid focusing of microscopic imaging system. The more accurate mathematical model is established to study nonlinear control method, improve the accuracy of magnetic shape memory alloy, The magnetic shape memory alloy driver can be applied to realize fast and high-precision focusing in microscopic imaging system, which provides a new idea for focusing device in microscopic imaging system.

In recent years, adversarial attacks have attracted more and more concentration, and various attack methods have made them a significant threat to deep neural network (DNN) security. However, most of the current research focuses on the attack algorithm without paying awareness to the critical information of the image itself. This paper proposes ANIS, a flexible, attention-guided noise injection system. The system combines image processing techniques with adversarial examples to inject different types of noise into areas of varying importance. It is an architecture that integrates easily with other attack algorithms. After testing, we confirmed that ANIS could improve the success rate of attacks, making it more difficult for defense systems to defend. We also compared the performance of different adversarial attack algorithms combined with ANIS and found FGSM with the most noticeable improvement in attack effect.

In order to reduce the workload of healthcare workers during surgery and free them from the task of recognising and delivering surgical tools, this project designs a vision-based robotic arm scalpel sorting system. To improve the accuracy and recognition speed, this design proposes an optimisation approach for YOLOv4-tiny recognition, which introduces a depth-separable convolutional lightweight backbone network in the feature extraction stage to reduce the computational volume of the model, which improves the FPS by 3.3%, the mAP by 4.3%, and the size of the model is reduced by41%compared to YOLOv4-tiny. In the experimental test, this design can automatically identify different scalpels, the success rate of surgical tool identification can be up to 95%, and the success rate of grasping can be up to 94%. This intelligent recognition system has the characteristics of higher growth nature, better reliability and application of medical devices, which can be used as a reference for research instead of complex medical and surgical environments.

We propose a novel snow removal algorithm. First, we make full use of the high-dimensional character- istics of video data and propose a tensor dimensionality reduction-based robust principal component analysis (TDR-RPCA) method to reconstruct clean background, which improves the traditional RPCA’s ability to handle four-dimensional data and solves the problems with artifacts residue. Then a penalty regularization term-based outlier detection method is used to detect moving objects and glob- al large snowflakes. A binary mask containing moving object and snowflakes is obtained. According to the mask, we design an adaptive thresholding strategy to remove snowflakes. Our scheme can effectively remove global snowflakes. Considering the sparse characteristic of snowflakes, tinny snowflakes in front of moving objects can be regarded as noise; hence, we deal with these tiny snowflakes using Gaussian-scale mixture-based simultaneous sparse coding. Experiments show that our method can remove snowflakes in videos captured during heavy snow conditions and eliminate artifact residue, which outperforms other state-of-the-art methods.

Naked barley is the fourth largest cereal crop in the world, and its output in Tibet, China accounts for 70% of the country’s total. In China’s highland barley-producing areas, the phenomenon of mixed planting and mixed harvesting of naked barley is common, which makes the refined management of the naked barley industry difficult, thereby reducing the economic value of specialty naked barley. This study proposes a new lightweight network model—RepSeed. RepSeed is based on the depthwise separable convolution and uses a multi-branch structure during the training phase to obtain more detailed features of the image to improve accuracy. During the inference phase, the multi-branch structure is changed into a single-path structure, thereby saving memory space and accelerating the model’s inference speed. Experiments were conducted using a data set containing 9756 images of naked barley seeds. The results show that RepSeed’s TOP-1 accuracy on the data set reached 90.40%, the parameter amount was 0.75M, the FLOPs were 114.6M, the latency was 23 ms, and the GPU memory space occupied was 181MB, the throughput in the GPU environment is 17074im/s.

This paper focuses on how SLAM can be improved to adapt to the harsh conditions of underwater environments. Compared with the common terrestrial environments, the use of visual SLAM systems in underwater environments will face the problems of light attenuation scattering and low visibility. In order to solve the problems of uneven light and blurred images faced by visual SLAM systems in underwater environments, we improved the visual part of the system by de-blurring the camera image to reduce the interference caused by underwater suspended objects. Histogram equalization is performed on the processed images to improve the visual feature point extraction by increasing the image contrast and reducing the effect of uneven lighting on the feature checker.

To provide a solution to the copyright protection problem of color images, a blind color image watermarking algorithm in the quaternion domain based on QR decomposition is proposed. Firstly, the Contourlet transform are applied to the R, G and B channels of the carrier image respectively, the low-frequency subbands are taken to construct a quaternion matrix, and the quaternion discrete cosine transform is applied to the quaternion matrix, then the inverse discrete cosine transform is performed on the real part matrix after the quaternion discrete cosine transform to get the intermediate process matrix. Finally, the above intermediate process matrix is split into non-overlapping blocks, the encrypted binary watermark is embedded into the upper triangular matrix obtained from the QR decomposition of non-overlapping blocks. The experimental date demonstrate that the watermark images extracted by the algorithm are still discernible after some common image processing attacks, so it has certain robustness.

The Segment Anything Model (SAM) stands as a foundational model for general-purpose image segmentation, showing promising performance in various natural image segmentation tasks. However, its inherent 2D characteristics often lead to inadequacies in capturing and integrating crucial information in 3D space, particularly when dealing with volumetric medical image data. To address this problem, we propose 3D ASAM, a novel approach for volumetric medical image segmentation. Specifically, our model feeds volumetric images directly into the SAM and CNN branches to help the model better understand the complex structure and texture of the images. We introduce a series of innovative 3D adapters in the Transformer block of the image encoder, while retaining most of the pre-trained weights in SAM. We utilize a parametrically efficient fine-tuning strategy and propose a new boundary difference loss function, aiming at improving segmentation accuracy and boundaries for better medical image segmentation. We conducted extensive experiments on pancreas tumor medical image datasets, and our method achieved competitive results compared with other advanced 3D methods when only single-point cues were provided.

The 3D human pose estimation algorithm achieves very high detection accuracy on standard datasets. However, the challenge of reducing model parameters and computational load without compromising accuracy remains a central focus of ongoing research. To address this problem, we propose a two-stage lightweight 3D human pose estimation framework, KPLitePose, which is characterized by a minimal number of parameters, reduced computational requirements, high accuracy, and fast convergence speed. In the backbone network of this framework, we designed KPLiteNet, a lightweight feature extraction network, which performs well in feature extraction while maintaining the lightweight nature of the model. Finally, a multi-view key point fusion method is proposed to reduce the impact of 2D key point noise on the results. We evaluated the performance of the framework on the Campus and Shelf datasets, achieving an inference speed of 65FPS on the Shelf dataset with a PCP3D score of 97.2%. Our framework cleverly balances inference speed and detection accuracy.

This paper presents the introduction of an innovative chaotic system termed ECNCM, which is based on two improved chaos models, improves the chaos of the model, and uses it as a new PRNG. Consequently, a novel image encryption technique has been crafted. The initial step involves generating a primary key, which is associated with the plaintext, using the SHA-256 hash of the plaintext image and additional relevant manipulations. Then a chaos matrix is generated based on the key, and the plaintext is encrypted through an index mapping algorithm. The image is scrambled for pixels, and finally the ciphertext image is obtained through encryption operations based on the chaotic array generated by ECNCM. The security of the algorithm is improved through the substitution-permutation network. The algorithm was tested and analyzed based on experiments. The experimental outcomes demonstrate that the algorithm provides notable effectiveness in encryption and strong security attributes, successfully protecting the data against common types of assaults, such as chosen plaintext attacks and noise injection attacks.

A thorough diagnosis of depression, a mental illness that affects many people worldwide, must be made in light of the patient's medical background, present symptoms, and pertinent examination findings. In recent years, scholars have increasingly favored machine learning algorithms for depression diagnosis or prediction models. However, achieving higher prediction accuracy in prediction models remains challenging when relying solely on one modality. This research suggests a multi-modal feature fusion model for predicting depression tendency that is based on bi-directional LSTM and vision transformer (ViT). The model demonstrates an accuracy of 70.00%, surpassing that of single-modal depression tendency prediction models and presenting a novel approach to depression tendency prediction.

The role of session-based recommendation in daily life is essential. There has been a focus on utilizing graph neural networks (GNNs) to recommend new items in the session contexts. However, existing methods only consider item transitions within individual sessions, posing challenges for GNNs in learning item embeddings and user preferences. To investigate the users’ intent and enhance the users’ satisfaction with new recommended items simultaneously, this paper proposes a novel dual-intent graph neural network approach (DGNN). Specifically, DGNN shall first incorporates not only pairwise item transition information within individual sessions but also the global information from the entire data set when adopting GNNs to learn item embeddings. Second, with item embeddings and items’ taxonomy tree embeddings fed into the dual-intent nets layer, DGNN synthesizes users’ intent by leveraging item information from the perspectives of both soft-attention and data distribution. Third, this paper has designed a novel method to infer the embeddings of new items. Experimental results on several data sets have verified the effectiveness of the proposed DGNN.

Computed tomography (CT) technology is widely used, but the X-ray radiation it emits is a concern. As a result, more and more research is focusing on how to maintain the quality of CT images while reducing the X-ray dose. With the advancement of deep learning technology, convolutional neural networks (CNNs) have been extensively applied in the field of CT reconstruction. However, CNNs focus only on local information and do not adequately consider the overall structure of the image. To address this issue, we have introduced an attention mechanism into the convolutional neural network to denoise low-dose CT images. We tested and evaluated the proposed denoising method using the AAPM-Mayo Clinic low-dose CT dataset. The experimental results show that our method can effectively remove stripe artifacts in LDCT images, preserve image details to a certain extent, and improve several metrics such as PSNR, SSIM, and RMSE compared to LDCT images.

Two main elements for the implementation of autonomous navigation in smart mobile devices are investigated. One is the simultaneous modeling and mapping of the environment, and the other is the autonomous navigation algorithm. Compared with ordinary mobile robots, track robots need both accuracy and speed. Through the study of SLAM technology and autonomous navigation algorithm, this paper proposes a set of autonomous navigation optimization scheme based on Teb algorithm, which realizes the rapid recognition and processing of target objects and road environment during the high-speed driving of intelligent mobile devices in the track, and ensures the safe, stable and fast driving of vehicles. This method improves and integrates local navigation algorithm and global navigation algorithm, and the experimental results prove that it has obvious improvement for mobile intelligent devices in indoor navigation, which has very important theoretical research value and practical application significance.

With the advancement of industrial automation, crewless operation has become a primary requirement for achieving "dark factory" requirements. This paper introduces a video-based understanding algorithm for crane hoisting status detection within industrial settings. An I3D-ResNet network enhanced with optical flow features is employed to classify the hoisting status. This algorithmic process can precisely capture and understand the dynamic changes during the hoisting process, significantly improving event recognition accuracy. Compared to using the I3D-ResNet network alone, the optical flow-enhanced network has demonstrated through experimental results its effectiveness in accurately recognizing hoisting status, providing robust support for achieving crewless production in industrial environments in the near future.

In the era of advancing deep learning, person re-identification has gained widespread application in domains such as video and security monitoring. Person re-identification seeks to recognize and match target persons across images taken by diverse cameras, thereby verifying whether the pedestrian subjects observed by cameras at various locations and times are indeed the same individual. This article categorizes current research into two main types: image-based person reidentification and video-based person re-identification. It provides a thorough overview and analysis of existing literature concerning classification methods, verification models, attention mechanisms, and metric learning, all centered around the research subjects. Additionally, it outlines the progression of datasets and provides insights into the anticipated future trends in person re-identification.

This study explored the morphological classification problem of the left atrial appendage. Given the diversity and unequal distribution of left atrial appendage categories, we propose a deep learning network based on attention mechanism. As the source of cardioembolic stroke and thromboembolism, occlusion of left atrial appendage is an important therapeutic approach. Therefore, accurate classify-cation of left atrial appendage morphology is crucial for the success of the surgery. By incorporating the attention mechanism, the model can effectively concentrate on image features, leading to improved classification performance and mitigating class imbalance concerns. Experimental results show that this method outperforms existing techniques in classifying left atrial appendage morphology, achieving an accuracy of 0.6584. It outperforms various classification networks and performs well in handling imbalanced data categories. The research findings have clinical significance for preoperative planning and postoperative recovery of left atrial appendage occlusion surgery.

Federated learning (FL) has become a cornerstone in medical image analysis, addressing the dual challenges of data volume growth and privacy concerns in centralized systems. Alongside FL, federated unlearning (FU) has emerged, focusing on the removal of sensitive information from models to comply with privacy regulations and ethical considerations. This paper introduces an efficient unlearning process using gradient ascent, aimed at balancing the “right to be forgotten” with the integrity of the learning process. Our approach, evaluated through the lenses of under-unlearning and over-unlearning, promises enhanced scalability and effectiveness in privacy-preserving machine learning, particularly in sensitive applications like medical image analysis.

The conventional approach for semantic segmentation of remote sensing imagery using encoder-decoder convolutional neural networks relies on the output of prior feature maps sequentially without considering the interactions between neighboring contextual feature maps with multiple resolutions. While the standard HRNet proposal has successfully improved multi-resolution semantic and spatial features to address the aforementioned issues, its lack of emphasis on boundary perception often results in inadequate target segmentation. Furthermore, a frequent occurrence of multiresolution contextual interaction in HRNet leads to the addition of a significant quantity of redundant information and amplifies the complexity of the model. Hence, to tackle the abovementioned issues, we propose a semantic segmentation network identified as HR-BMNet, which incorporates boundary sensitivity and multiple-resolution learning. The idea associated with standard HRNet is adopted as the foundational architecture. We extend novel boundary perception and multi-resolution fusion attention modules, integrating channel attention mechanisms. The strategy provides an ex-tensive optimization of edges and the efficient capture of crucial multi-scale features. During the feature combination stage, the boundary insights are employed to augment the semantic information, thereby mitigating the spatial details loss, enhancing the intra-class semantic consistency, and achieving superior segmentation. The efficacy of the proposed method is validated through comparison and ablation experiments conducted on the ISPRS Vaihingen and CSRSD datasets. Among the experiments conducted, the best ones attained a mean Intersection over Union (mIoU) of 72.11% on the Vaihingen dataset and 89.28% on the CSRSD dataset, respectively.