This PDF file contains the front matter associated with SPIE Proceedings Volume 13210, including the Title Page, Copyright information, Table of Contents, and Conference Committee information.”

In the industrial control domain, accurate network traffic prediction is crucial for optimizing resource configuration and enhancing system stability. Addressing the limitations of existing methods in handling both nonlinear long-term dependencies and short-term linear features simultaneously, our proposed model adopts a two-fold approach. First, we decompose the traffic sequence into distinct components, each characterized by specific features. For nonlinear dependencies, we employ the Bi-LSTM model, capturing bidirectional relationships and incorporating a self-attention mechanism for adaptive feature weight adjustments. Simultaneously, for linear features, we leverage the NeuralProphet model. In summary, our model outperforms traditional time series prediction methods, demonstrating improved accuracy in forecasting overall network traffic and offering enhanced precision in predicting traffic burst modes and troughs.

The complexity of modern software has resulted in an increase in vulnerabilities and bugs. Identifying all vulnerabilities is challenging due to technological limitations, so patching bugs found during the usage phase is necessary. This leads to high maintenance costs. Furthermore, undisclosed vulnerabilities can be easily exploited by hackers. Static code analysis is a technique used to identify bugs and defects in a program during the development phase. This can greatly improve the reliability and quality of the code. However, it may have the disadvantage of low analysis efficiency. To tackle this issue, we present FCG, an algorithm for rapid call-path generation based on call graphs. FCG can swiftly identify all feasible function call sequences that lead to a particular location in a program, thereby enhancing efficiency for subsequent analysis. In this paper, we conduct an experimental comparison between the FCG algorithm and its closest analogue, DFS. The results demonstrate that FCG outperforms DFS in the presence of a large amount of code.

Convolutional neural networks have been widely used in recent years for image analysis in areas such as tumor segmentation and disease detection. However, a growing body of research has shown that applying small disturbances to an image can wreak havoc on the results quite differently from a correct diagnosis. This is of great concern to organizations that have already deployed AI medical detection systems, and related research is proceeding at a rapid pace. In this paper, we investigate different ways of attacking medical images against adversarial attacks, and the reasons why convolutional neural networks show performance degradation against perturbation noise are analyzed. We propose a new method to improve the robustness of the convolutional neural network classification system by adding attention modules at different locations of the model to improve the performance of the model's feature extraction as a way to improve robustness. Finally, the performance of the network model is examined using the most popular attacks available. We validate on a medical image dataset and demonstrate that the proposed method allows the convolutional neural network to successfully defend against attacks on counter-attack samples without significant performance degradation. Our study suggests that these findings may provide some positive effects on future medical image analysis and processing as well as exploring deep image features.

This paper briefly introduces the traditional design mode in the research institutes, based on the design efficiency improvement, computing resources optimization, localization of software and hardware applications, build the design based on private cloud integration platform, solve the research institutes design in the process of computing resources as a whole, localization alternative, system matching. Finally, suggestions are put forward on the focus and main expansion direction of scientific research institutes in the application design integration platform.

Recipe recommendations are highly sought after in today's society, however, accurate personalized recommendations remain a challenge. Our research focuses on improving traditional recipe recommendation by introducing heterogeneous graph neural networks to mine the information of higher-order co-signals, such as ingredients and tastes, in order to improve recommendation accuracy. First, by constructing a recipe information graph, we achieve effective integration of multi-layered information such as users, recipes and ingredients. In addition, we combine the flavor information to find out the cuisine preferences of different groups of people and identify potential audience groups. Finally, the model effectiveness is improved by self-supervised learning of graphs. Experimental results on a large-scale food recommendation dataset show that the method improves over the optimal baseline method in four metrics, namely, Precision@10, HitRate@10, NDCG@10, and MAP@10, thus demonstrating the effectiveness of the method.

Introduced a multi target fusion and tracking method based on dual base coherent MIMO radar. The synchronization of the crystal oscillator and local oscillator signals of the dual base radar is achieved through fiber optic transmission, thereby achieving coherence between the dual base radars. By using the ESPRIT algorithm to estimate the transmitting azimuth DoD and receiving azimuth DoA, a joint estimation result of the target position by the dual base radar is obtained. By matching intermediate distance variables, the calculation results of the target position by the receiving radar are obtained. Match the target position calculation results of the transmitting and receiving radars with the joint estimation results of the dual base radars to obtain the vector velocity of the target. Perform multidimensional Kalman filtering based on the multi-dimensional position results and vector velocity of the target to achieve target tracking. This method can significantly improve the accuracy of target fusion and tracking, and shorten the time from target detection to target tracking.

Recently, Information and Cyber Security are taken more seriously. Traditional security technologies cannot deal with issues and problems in the new era completely and correctly. So, advanced technologies, such as Artificial Intelligence (AI) and interrelated technologies are becoming promising and successful additions based on the rapid development of Computer Science. This paper aims to introduce the research status of applications of advanced technologies in the fields of Cyber Security based on bibliometrics and text mining, including some limitations and challenges. We survey the relevant literature downloaded from Web of Science, comprising papers mainly published in between 2010 and 2022. Many results concerning the focused areas and development tendency, etc. are presented in this paper.

The motion blur of the star has spatial non-uniformity in the image plane of the star sensor. Therefore, the star simulator simulation model must consider the effects of motion blur and its spatial non-uniformity to provide a more realistic and complete test for the star sensor. During an exposure period, the trajectory of the deformed and blurred star spots in the image plane is an arc-shaped curve. First, the length of the motion blur can be decomposed into radial and normal lengths, which can be represented by the position coordinates of the star and the corresponding to the angular velocity component. An expression for the star image-motion is derived in terms of the exposure time, star location, vehicle attitude angle and the focal length. Second, based on the energy distribution characteristics of moving stars, three methods of simulating blurred star spots were analyzed. Finally, the stars map of spatial non-uniformity of the motion blur was simulated.

In this paper, we propose a novel deep learning network model for stereo matching that achieves higher accuracy and better real-time performance. We achieve the effect of streamlining the complexity of the network model by using the lightweight network model MobileNetV3 for feature extraction. By using feature maps at four different scales, combined with the attention feature blocks in the network, the network can combine contextual information with better sensing ability, and the network model achieves a more accurate and efficient processing effect through multiscale feature fusion. We pretrained the model on the SceneFlow virtual datasets and fine-tuned it with multiple epochs of training on the KITTI datasets, which resulted in better generalization of the stereo matching model and more accurate disparity prediction results for real traffic and road environments. Compared with some existing network models such as PSMNet, GANet, our network achieves a faster processing rate with guaranteed accuracy, which meets the real-time requirements of network models under general conditions.

This paper presents a novel software development methodology for building a low-code development platform core architecture by using abstract syntax trees. The aim is to strike a balance between development efficiency and reducing technical barriers while maintaining high scalability. The paper introduce the basic principles of low-code development architecture based on AST and elaborating on the relationship model of three crucial user groups from the bottom support to the business construction. On the other hand, a general-purpose program language AST construction algorithm is proposed, which demonstrates better performance and lower memory consumption in handling large-scale multilingual business scenarios.

In order to make the lightweight network more effective to understand the feature representation of no-reference image quality assessment (NR-IQA). We propose a dual-stream SqueezeNet combined with gradient maps. SqueezeNet can improve the accuracy of the training model while reducing the number of parameters. The dual-stream includes two input channels: RGB image patches and gradient images. RGB image stream pays more attention to the intensity of image information, and gradient stream focuses on extracting detailed image feature information in detail. We use SqueezeNet to train image patches and gradient images separately, and finally obtain the final image quality score by averaging the scores of the patches. We conducted extensive experiments on the basic data sets LIVE, TID2013, and CSIQ. In the experiment, our proposed approach achieves an SROCC score of 0.895 with only 0.42 M, in the WN distortion type of LIVE, the highest SROCC can reach 0.991, the JPEG distortion type of CSIQ can reach 0.966. The cross-data set verification experiment proves that the proposed model has high robustness.

To solve the problem that the ground target in SAR image is small in scale and there may be complex background around it, a improved SSD model for ground target recognition is proposed to ameliorate the accuracy and the speed of detection about SAR small target. The feature fusion module is added to the SSD model, while the context information is introduced to the large-scale feature maps. The attention mechanism is introduced, when the squeeze-excitation module with lower computational cost is added, so that the model can reconstruct the every channel. The experimental results on the MSTAR dataset suggest that the improved SSD is superior to other typical one-stage algorithms in accuracy and speed of detection, which can satisfy the needs of real-time ground target recognition.

Deep learning for speech enhancement has become a popular trend. With the introduction of attention mechanisms, the field of speech enhancement has seen further advancements. In previous research, Transformer is commonly employed as the core network for speech enhancement due to its attention capabilities, but it has certain limitations in extracting local features. To address these issues, we propose a novel speech enhancement model named CBSENet. This model utilizes the CRN architecture and adopts Conformer as its core network. The Conformer's capabilities in global and local feature extraction have been validated. The CBSENet model demonstrates excellent performance in speech enhancement tasks and outperforms the baseline network in both objective and subjective evaluation metrics.

Based on the background of underwater acoustic communication, this paper first examines the Orthogonal Time Frequency Space (OTFS) modulation technique, which exhibits strong time delay and Doppler tolerance. This modulation technique effectively mitigates time-selective fading caused by Doppler effects and offers superior performance advantages. Subsequently, the study delves into an analysis of current symbol detection algorithms in channel transmission, focusing on the limitations of the message passing (MP) algorithm. The paper proposes specific enhancements to the MP algorithm, based on its original framework, and assesses the system's performance using different modulation techniques, path numbers, and Doppler factors through simulation. The findings indicate that OTFS modulation outperforms OFDM modulation, and the improved MP algorithm surpasses Doppler modulation. Furthermore, the performance of the enhanced MP algorithm exceeds that of OFDM modulation, demonstrating its superiority over the traditional MP algorithm.

Efficient super-resolution networks have been more well-known in recent years due to their effectiveness in increasing image resolution with the least amount of processing cost. This research introduces a novel lightweight super-resolution method that prioritizes efficiency by integrating channel information, fusing spatial elements, and paying attention to multiscale characteristics. The suggested network makes use of cutting-edge deep learning methods, concentrating on the extraction of spatial characteristics to maintain fine details when upscaling. Furthermore, multi-scale characteristics are carefully integrated to handle various image structure levels. The channel information fusion improves the network's capacity to identify intricate linkages in the data. The suggested method achieves significant gains in image quality and resolution, as shown by the experimental findings, demonstrating the usefulness of the lightweight design in practical applications. The results we have shown highlight our network's ability to achieve a higher super-resolution performance while maintaining computational efficiency, which makes it a viable option for real-world picture enhancing applications.

Currently, semi-supervised learning methods have become popular in medical image processing tasks, this paper proposes a medical image classification method based on a semi-supervised pseudo-label augmented generative adversarial network. This method combines the generative adversarial network and pseudo-label generation technology, introduces the FID evaluation index to train the generator with good performance for the problem of poor classification performance of GAN, and uses only a small number of real images to generate a sufficiently large number of pseudo-images for model training, which is combined with the pseudo-label generation method of teacher-student network introduced based on MixMatch to categorize the medical images. Through experimental validation, this paper's method outperforms other semi-supervised classification methods in terms of classification accuracy with limited labeling and achieves classification on three publicly available medical image datasets. In addition, we conducted ablation experiments to verify the algorithm's effectiveness.

To address the issue of low recognition capability and weak generalization of traditional synthetic aperture radar (SAR) target recognition methods under the condition of limited training samples, a Few-shot SAR target recognition based on Rectified Gaussian Prototypical Networks is proposed. This method utilizes an encoder to obtain the embedding vectors of the images, where some components of the embedding vectors represent confidence region estimates around the embedding points and are expressed as Gaussian covariance matrices. Subsequently, initial prototypes are constructed based on them. Then, the prototypes are further refined through the Bias Decrease module. Finally, the target category is determined by computing the cosine similarity between the test sample and the prototypes. The proposed method is tested and validated on the MSTAR dataset. The results show that compared to the nearly ineffective performance of traditional neural networks under small sample conditions, the proposed method can achieve more robust feature extraction and recognition. Using a prototype network as a benchmark, there is an improvement of more than 10% in recognition rate. Meanwhile, when compared with various classic few-shot algorithms, the proposed method has achieved better results, demonstrating a certain degree of superiority.

This article briefly introduces GPFS and GFS shared file systems, discusses the role of GPFS and GFS shared file systems in the field of server management, and compares the GPFS and GSP shared file systems of the PDM server cluster in the research institute; After testing, it was found that the GFS shared file system is more suitable for PDM server clusters. Three running servers in the PDM server cluster can concurrently access files in the GFS shared file system; Enable all three servers to independently provide services to the outside world. Moreover, any failure of two servers will not cause the PDM service to stop, ensuring the continuous progress of research and development in the research institute.

The explosive increase in cloud traffic has put forward higher requirements for network performance. As a key technology, smart network cards (SmartNICs) are of great significance for solving network performance bottlenecks in cloud environments. However, there is currently a lack of effective monitoring methods for the traffic flowing through smart network cards, making it a black box, which is not friendly to ensure the quality of service(QoS) for users. To alleviate it, in this paper, we propose a low-overhead collaborative design for data collection based on XDP in which the XDP program is attached to the VF representor(VF rep) which is the mandatory path for every traffic flow that is about to be offloaded to the SmartNIC to capture the necessary traffic information, and then the OVS command is used to obtain detailed traffic statistics offloaded. The collaborative design not only provides sufficient valuable information for SREs to understand the health status of the network, which is important for the QoS assurance but also has negligible performance overhead due to the inherent low overhead of the XDP program. Experimental results show that the proposed method costs less than 1% latency and throughput overhead to provide the desired information, which demonstrates our method is effective.

Heterogeneous Graph Neural Networks (HGNNs) have emerged as powerful tools for handling heterogeneous graphs. However, current HGNNs often rely on meta-paths or intricate aggregation operations. In response, we introduce a heterogeneous graph neural network based on dual-view graph structure augmentation, which consists of three aggregation processes. By leveraging both node feature information and graph topology structure information, our method selecting homogeneous neighbors for nodes and constructing homogeneous views. Subsequently, it learns node representations through aggregation on these views and the original graph. Through extensive experiments on three widely used real-world heterogeneous graphs, our method demonstrates its simplicity and effectiveness, and outperforms the most of existing models in the task of heterogeneous graph node classification.

As an important part of blockchain technology, smart contracts have attracted strong interest from industry and academia. They provide the foundation for implementing various blockchain applications and play a key role in the blockchain ecosystem. However, the frequent occurrence of smart contract vulnerabilities has resulted in significant economic losses and serious damage to the blockchain-based credit system. Currently, the security and reliability of smart contracts have become an emerging research field. Recently, deep learning methods have achieved certain results in mitigating the vulnerability problem of smart contracts, with the BERT model being widely used due to its good performance. However, the existing BERT model solely relies on features extracted from the last layer, leading to incomplete classification features. To address this issue, we propose a multi-layer feature fusion model that can accurately identify smart contract vulnerabilities. Our fusion model integrates features from multiple layers of the BERT model and employs a series of fusion strategies to enhance the comprehensiveness and accuracy of the extracted features. We extensively tested and verified the effectiveness and performance advantages of our proposed multi-layer feature fusion model. Compared to traditional single-layer feature extraction methods, our model demonstrates higher accuracy and a lower false positive rate in identifying smart contract vulnerabilities.

The scale and functionality of modern embedded system software is expanding, and the ponderous network data traffic arises. To ensure the smooth operation of the system, network traffic monitoring is essential. The article studies the application of deep learning technology in the field of network traffic monitoring and analyzes the use of deep learning for traffic classification, traffic prediction, fault management, and security. It proposes to apply deep learning in the field of embedded network traffic monitoring, analyzes the challenges, introduces the existing research results, and looks forward to the new methodology, which is of great significance to deep learning as well as the embedded field.

This study examines the impact of varying wind and wave levels on the radar cross section (RCS) analysis of ship targets at different sea surfaces. A ship model was established and the far-field single station RCS of ship targets was analyzed under various sea conditions using a hybrid algorithm of physical optics and moment method (PO-MOM). The study utilized FEKO software to simulate and calculate the far-field single station RCS of the European Buffalo class hovercraft, while also improving the hovercraft model to account for seawater. Additionally, the study analyzed and studied the target RCS of the ship under different sea conditions using the PO-MOM hybrid algorithm, and ultimately utilized the CSI algorithm to compensate for it. The findings indicate that as the sea level rises, the ship's RCS decreases.

This paper presents a UAV visual simulation platform based on 3D holographic display technology, designed for UAV simulation training and real-time dynamic simulation of large scenes. The platform involves key modules such as UNIGINE real-time 3D engine, UAV flight simulation, optoelectronic video simulation, SAR image simulation, multispectral image simulation and computational holographic 3D display. The simulation platform has realized the 3D holographic display of UAV flight trajectory, providing immersive simulation training for operators to control the UAV to perform takeoff, flight, reconnaissance and landing tasks. The application requirements of the UAV system have been satisfied, and it has important practical value.

Deep graph clustering is a challenging field that help us analyze complex network data which involves dividing nodes into non-overlapping clusters. A key issue in this field is the reliance on complex data augmentation. Traditional data augmentation methods not only increase data processing time, but also require specialized knowledge to determine appropriate strategies. At the same time, traditional data augmentation may not fully utilize the topological properties of graph data, thus limiting the potential of graph clustering models. Based on these, we propose a novel unsupervised learning approach for automatically identifying and implementing effective data augmentation strategies, focusing on community-based augmentation. This approach dynamically analyzes graph data structures, reduces manual intervention, and improves the efficiency and accuracy of deep graph clustering. Our experimental results validate the efficiency of our method, which outperforms the models without considering information enhancement.

The performance monitoring of gas turbine is an important global issue that is related to the stable and safe operation of gas turbine equipment. With the development of gas turbines, the complexity of its structure and the harsh conditions of high temperature and high pressure have brought many difficulties to measure. To a certain extent, it is impossible to measure even at the exit of the burning room or the high -voltage turbine inlet, which leads to how to use these available parameters to ensure the health state of the gas turbine. The emergence of digital twin technology provides a feasible way for the health monitoring of gas turbine, which has set off a wave in the field of gas turbine's health monitoring. Digital twin is a numerical copy based on physical data. It can truly reflect effective information and have the ability to self -correct and predict the future. This article focuses on the logical framework of the performance monitoring operation of the gas turbine, builds the calculation module and optimization function of the functional component of the gas turbine, and verified it through experiments. The results show that the system can be used for the construction of the digital twin model of the gas turbine, which has certain engineering significance for future performance monitoring.

Wireless Battery Management System (WBMS) is a solution that uses wireless communication technology to monitor the battery pack of electric vehicles. Compared to traditional wired battery management systems (BMS), wireless transmission technology can provide a more efficient, flexible and scalable battery management solution. This paper presents a reliable WBMS communication protocol that utilizes a dual-cycle frame structure to support the sampling data transmission of both battery pack voltage and bus current. A WBMS network simulator is developed based on OMNeT++ framework, and the network performance is evaluated with different propagation parameters. The simulation results show that the proposed wireless communication protocol has remarkable performance even in the presence of WiFi and Bluetooth interference.

In the field of source-based software defect prediction, it is necessary to convert the source code into the data form that the model can process, which is called word embedding. The commonly used word embedding models are Word2Vec and BERT models. Meanwhile, in software defect prediction, accurate feature representation is crucial to the performance of the defect prediction model. However, feature redundancy problems in source code, such as highly similar word vectors caused by code appearing in pairs, result in a set of word vectors generated when training word embedding models that may degrade model performance when applied to defect prediction tasks. In order to alleviate this problem, a CodeSift method is proposed in this paper. By calculating the similarity of each pair of word vectors in the code word vectors generated by the word embedding model and interpolating the generated code word vectors, the problem of feature redundancy between word vectors is reduced. CodeSift generates a new word vector from a highly similar word vector, and retains the original information by assigning weights, thus improving the compactness and information richness of feature representation. Experiments show that the F1 value of the defect prediction model is improved by using CodeSift method, and the false positive rate is lower than that of the original model.

Fusion of texture information and semantic information from visible and infrared images is important for many fields. However, existing algorithms extract features with a limited range of receptive fields which causes the loss of contextual information. This paper introduces DCTDFusion, a two-branch network integrating dilated convolution and Transformer to address this issue. We propose Residual Dilated Convolutions Block (RDCB) to extract local information of different receptive field ranges. To get the global information at multiple scales, we introduce the Transformer Dilated Convolutions Block. In addition, the Sobel feature extractor is used to retain the gradient information of both branches. Numerous experimental evidences show that the fused images of DCDTFusion contain rich information of significant targets and background in source images. Additionally, our method improves about 20%, 4%, 1% and 2% in MI, VIF, QAB/F and SSIM metrics comparing with the second best.

Many programmers will look for available software in the code base, but there are lots of software without classification labels in the code base. If the software is not classified correctly, it will prevent programmers from retrieving the libraries they need, affecting the development process. Based on the above background, this article selects the Maven code warehouse as the research object, using the convolutional neural network model (CNN) and long short-term memory network model (LSTM) in deep learning. This paper uses software code, text, and dependencies as input to the model. This paper aims to assess the accuracy of software classification. At the same time, the parameter configuration of the model is continuously optimized during the training process to find the parameters that enable the model to achieve optimal results. The experimental results show our model can help developers find suitable software and can be useful in classifying software systems in open-source repositories.

Current software evaluation work based on complex networks rarely considers the complexity of nodes themselves and the multiple coupling between nodes, making it difficult to accurately identify high complexity and high coupling modules in software. This paper proposes a method for evaluating software node complexity in complex functional granularity networks. By analyzing the static structure of complex software and clarifying the different types of dependencies between functions in the software, a function level complex coupling network is constructed. Propose a node risk importance index, considering the criticality of nodes in the network and the complexity of nodes themselves, to identify highly complex modules in software. Experimental results demonstrate that the NRI method proposed in this paper can more effectively identify complex nodes in software.

The rapid development of digital technologies, such as big data and artificial intelligence, has led to the emergence of Internet-based platforms for government services and credit dissemination. These platforms have accumulated a vast amount of credit data. However, the current social credit system still faces issues such as opaque credit information, easily tampered data, and insufficient traceability. Our proposal uses blockchain technology to establish a digital government-based credit data-sharing framework (TCDSF). This framework provides a tamper-proof trust mechanism that enhances information transparency, reduces the cost of inter-organisational data sharing, and mitigates data security risks; To protect the credit subject's privacy during the trust collaboration process, we have designed a master-slave chain sharing and collaboration mechanism. This mechanism uses both on-chain and off-chain storage modes to reduce the storage pressure of distributed nodes and enhance flexibility and scalability; We ensure the security of shared data in the channel through role-based access control measures and an asymmetric encryption mechanism for credit data. Experiments have demonstrated that the framework is capable of providing continuous, efficient, and reliable management and traceability services, as well as efficient resource utilization and adaptable data processing and storage, compared to a single chain.

Invoice data is critical and highly sensitive for enterprises, yet it stands as one of the most extensively used data categories in the financial credit sector. Participants in the credit market, engaged in corporate lending activities, face the dual challenge of ensuring the security of enterprise invoice data and safeguarding the integrity of institutional rules. Traditional data computation and encryption present a dilemma, as two encrypted datasets cannot directly engage in calculations, and two datasets participating in calculations must be in plaintext. This necessitates the convergence of enterprise data and institutional rules for computation. Consequently, two sets of data must be exposed to one of the parties or a mutually trusted third party. In practice, identifying a genuinely trusted third party proves to be a formidable task. This paper introduces a federated privacy computing method based on an oblivious transfer algorithm in the context of risk control admission rules within the credit risk management field. This approach enables both parties holding data to determine admission rules without disclosing their own data.

Digital twin technology is being applied in various fields, particularly in Industrial Internet of Things. Fault detection and predictive maintenance function is an important application of digital twin. Anomaly detection of time series data collected by various sensors is one of the main technical means to realize predictive maintenance and fault detection function. At present, anomaly detection in digital twin systems faces some problems: the lack of abnormally labeled data, the difficulty in capturing the correlation and temporal dependence between sensor data, and the difficulty in processing multi-variable and complex sensor data. With the development of data processing technology and artificial intelligence, various new algorithms for time series anomaly detection have emerged in an endless stream, which brings new technical solutions to the anomaly detection of digital twin systems. This paper presents a deep learning-based anomaly detection solution for digital twin systems, offering new perspectives and advanced methods for enhancing anomaly detection in digital twin systems.

An important direction for future 6G mobile communication networks is achieving global wireless coverage. To realize comprehensive access services, research has been conducted on multi-satellite cooperative computation offloading and resource allocation issues based on inter-satellite links (ISL). Ground users adopt a partial offloading scheme, aiming to minimize the weighted sum of ground user latency and system energy consumption, and an optimization problem is established. A joint optimization algorithm for low earth orbit (LEO) satellite task offloading and resource allocation based on the Multi-Agent Proximal Policy Optimization (MAPPO) framework is proposed. For the joint computing issue among LEO satellites, a Kuhn-Munkres algorithm (KM) based on a greedy algorithm for pre-allocation results followed by a secondary allocation has been designed. On the other hand, to accelerate the convergence speed of the reinforcement learning algorithm, an action masking mechanism is introduced. The action space of agents was optimized using the Particle Swarm Optimization algorithm (PSO), obtaining a prior distribution of the action space, thereby narrowing the search space of the algorithm. Simulation results indicate that, compared to the MAPPO method without the use of trajectory rehearsal action masks, the proposed method demonstrates approximately a 51.5% improvement in convergence performance and a 6.4% enhancement in convergence outcomes. Additionally, the average computational delay for tasks remains stable at around 0.32 seconds, which represents a reduction of up to 5.4% compared to PSO.

Generative Adversarial Networks (GANs)¹ have demonstrated remarkable success in generating realistic images, and are increasingly used for image-to-image translation tasks in medical imaging. However, methods like Pix2pix², which rely on pixel losses, require strict alignment of training samples. In medical MRI, strictly aligned training images are often difficult to obtain, as imaging of two modalities is not usually performed simultaneously. Although adversarial networks represented by CycleGAN³, which use consistency losses, can be trained with unpaired images, they often generate images that are biased towards low frequencies and are not clear enough, which may lead to the loss of important anatomical structures in the generated images. To address this issue, we propose a supervised GAN framework based on registration losses, AligFT GAN. The framework, based on deformation field theory, considers misaligned labels as noisy labels, models the noise with a deformation field, and trains the generator using an additional alignment network that estimates displacement vector fields to adaptively fit the misaligned noise distribution. In addition, AligFT GAN employs different types of frequency space losses to regulate the frequency content of the generated images. It combines the well-known properties of MRI K-space⁴ geometry to guide the network training process. By combining our method with GANs with cycle consistency losses, we can mitigate the impact of using misaligned data for training and frequency bias, to synthesize high-quality images.

To enable large-scale and efficient deployment of artificial intelligence (AI), the combination of AI and edge computing has spawned Edge Intelligence, which leverages the computing and communication capabilities of end devices and edge servers to process data closer to where it is generated. A key technology for edge intelligence is the privacy-protecting machine learning paradigm known as Federated Learning (FL), which enables data owners to train models without having to transfer raw data to third-party servers. However, FL networks are expected to involve thousands of heterogeneous distributed devices. As a result, communication efficiency remains a key bottleneck. To reduce node failures and device exits, a Hierarchical Federated Learning (HFL) framework is proposed, where a designated cluster leader supports the data owner through intermediate model aggregation. Therefore, based on the improvement of edge server resource utilization, this paper can effectively make up for the limitation of cache capacity. In order to mitigate the impact of soft clicks on the quality of user experience (QoE), the authors model the user QoE as a comprehensive system cost. To solve the formulaic problem, the authors propose a decentralized caching algorithm with federated deep reinforcement learning (DRL) and federated learning (FL), where multiple agents learn and make decisions independently.

Multi-constraint travel route planning is a challenging problem, which involves multiple constraints such as transportation, time, cost and traveler preference. In recent years, the development of deep learning technology has provided new possibilities to solve this problem. In this paper, a method combining deep learning and LKH algorithm is proposed to solve the multi-constraint tourism planning. First, we collect tourism data, including scenic spot information, transportation information and traveler preferences. Then, we trained the tourism data using deep learning models, such as neural networks, to learn the correlations between attractions and traveler preferences. We designed a multi-task learning model that can simultaneously predict multiple constraints such as travel time, distance, cost, etc. Finally, we apply the LKH algorithm to the output of the deep learning model to obtain the optimal travel route. Experimental results show that our method can efficiently generate travel routes that meet travelers' needs under multiple constraints. This study provides a new solution for multi- constraint travel planning, and also provides a practical application case for the combination of deep learning and traditional algorithms.

This paper proposes a method for super-resolution reconstruction of fruit images. Firstly, in order to eliminate background interference and reconstruct fruit regions more targeted, the K-means clustering algorithm based on Lab color space is used to automatically extract the fruit regions from whole fruit images. This method can be applied to fruit images with different backgrounds. Then, the ANR model is used to achieve fast reconstruction for the fruit region image. Finally, a global error compensation optimization algorithm is used to further correct global brightness or color errors in the fruit region image and improve its overall quality. The experimental results show that the algorithm proposed can accurately extract regions of interest for fruits in images with different complex backgrounds. Meanwhile, PSNR data indicates that the image reconstruction and optimization algorithm proposed has certain advantages and practicality.

Computer information processing technology meets this need and is used in practical work. The era of big data has also put forward new requirements for the information processing ability of computers, and previous computer information processing technology can no longer meet the development needs of the country and society. This article optimizes computer information processing technology based on big data. Through DM (Data Mining), the original information processing method has been transformed, providing new directions and development ideas for its development, continuously improving information processing efficiency, optimizing information processing structure, providing more convenient information services for people, and promoting the development of information processing technology. The experimental results show that the comparison method shows exponential growth, while the method in this paper shows almost linear growth, and the execution time remains at 7.6 seconds. This is because the comparison method uses serial execution, and the time complexity is exponential. This method optimizes the complexity to linear through parallel execution. This indicates that the missing information detection method proposed in this article significantly optimizes the detection execution time, thereby effectively improving the detection latency.

To achieve high-precision control at a lower cost, researchers propose combining the traditional machine arm method with new hardware, Long-short Term Memory (LSTM) Network, and Proportional Integral Derivative (PID) Control. The machine arm method provides reliable control but lacks precision. LSTM Networks capture long-term dependencies, improving accuracy. PID Control continuously adjusts control signals for stability. This integrated approach offers a cost-effective solution for industries seeking precise control.

Today, with efficient information dissemination and free interaction, digital+network has almost become the main channel for cultural transmission. The literary creation and dissemination of Chinese medicine culture has a relatively broad influence in modern society. This paper aimed to study how to analyze and study the popular science information of TCM (Traditional Chinese medicine) based on the block-chain operation of new media communication. It also described the new media block-chain information filtering method. This paper took the documentary “the Tale of Chinese Medicine” as an example to analyze its dissemination in the new media. Through the analysis of the gender and age distribution of the people concerned during the whole broadcast period, this paper found that the focus group was mainly 20~49 years old, reaching 96.00%, which was the main audience group of the new media communication of “the Tale of Chinese Medicine”. Among them, the audience aged 30-39 accounted for 51.00% of the total audience, which was closely related to social perspective and health preservation. The inevitability of TCM culture inheritance is due to the demand of TCM culture inheritance on the one hand, and the unique role of national cohesion on the other hand. Chinese medicine culture is a unique culture of China, which has special significance for the Chinese people of all ethnic groups.

This paper proposes an algorithm strategy that utilizes the empirical Bernstein inequality to reduce the number of random walk samples in the single pair algorithm, thereby lowering the algorithm's theoretical complexity. An algorithm named EBSim is introduced, which combines the empirical Bernstein inequality by iteratively adjusting the number of samples, and it is proven to have strict error guarantees. The paper also provides a detailed analysis comparing this algorithm with the Monte Carlo algorithm, demonstrating that our algorithm achieves tighter confidence intervals, lower algorithm complexity, and higher efficiency in handling the same query problems. Experimental results show that, compared to the widely used Monte Carlo query algorithm, EBSim significantly improves algorithm efficiency while ensuring a given error tolerance.

Most of current image steganography algorithms are founded on the principle of data embedding, the data embedding operation always inevitably lead to the distortion of the cover image and liable to catch suspicion of attackers. Aiming at this problem, an image steganography without data embedding utilizing style transfer and zero-watermarking is introduce in this paper for the security of secret information transmission. First, the sender makes Quaternion Polar Harmonic Fourier transfer on the original image to obtain Quaternion Polar Harmonic Fourier moments(QPHFMs), and conduct XOR operation between the secret information to be transmitted and these generated moments, resulting in the key watermark for confidential information transmission. Subsequently, the original cover image is modified by utilizing the style transfer network to obtain a stylized image, which would be transmitted over a public channel. On the receiver side, the elaborately designed de-stylized network is employed to reconstruct the original image. Then, the QPHFMs would be re-obtained with Quaternion Polar Harmonic Fourier transfer on the reconstructed image, and the secret information are extracted by XOR operation between the re-obtained QPHFMs of the reconstructed original image and the key watermarks. There is no embedding operation of secret information in the whole secret information transmission process. Extensive experimental results indicate that the proposed scheme can implement the task of secret transmission safely and successfully.

Based on the product tracing and quality monitoring requirements of Lijiang snowpeach, a solution using image matching for product identification was proposed. In response to the slow speed and false matching problems of traditional SIFT algorithm in snowpeach image matching, an improved SIFT algorithm was proposed, which uses Laplace algorithm for image enhancement, extracts more image information, adopts fixed-position feature vector extraction method, reduces the feature vector of image keypoints from 128 dimensions to 64 dimensions, and uses center-distance method to remove edge keypoints, reduces wrong matches. Experimental results indicate that the algorithm effectively improves the speed and accuracy of image matching.

With the development of network security technology, the situation of network intrusion is becoming increasingly serious. Traditional network intrusion detection methods have the problem of low detection efficiency and the lack of encryption processing in detection rules, which may be cracked by attackers. In order to improve the intrusion detection ability of frequent data transmission in the network, In this paper, a federated learning method based on blockchain technology is proposed, which uses blockchain technology, chain data structure and data block rules to verify the relevant information in passwords, and inserts password strings in multiple blocks to improve the intrusion detection ability of network information. At the same time, this paper analyzes the reasons of information theft and interference, summarizes the characteristics of network intrusion, puts forward the design idea of integrating blockchain technology, and verifies the correctness of this processing method, the fitting of network data and the degree of intrusion detection through actual cases. MATLAB simulation results show that the federated learning method has better network intrusion ability, the intrusion detection degree is over 90%, and the intrusion detection ability rises first and then declines, the overall index is better than the online monitoring method. Therefore, the fusion of blockchain and federated learning method proposed in this paper is suitable for the optimization of network intrusion monitoring. There are still some shortcomings in this research, mainly because there are few research materials related to blockchain technology, which leads to less research data in this paper, and more data will be collected for analysis in the future.

Efficient algorithms for generating digital camouflage patterns are widely used in the fields of smart windows, electronic skin, and bionic electronics. In this paper, a digital camouflage pattern generation algorithm based on an improved GA-Kmeans clustering algorithm was proposed. The proposed algorithm renders accurately extraction of the background domain colors while maintaining the background textures. The key features of our algorithm include: (1) a rectangle blocks segmentation and random scrambling algorithm which works fast and ensures maximum retention of local texture details is designed to process the neighborhoond background image and disorder the background textures, (2) an improved GAK- means clustering algorithm which works as a global search algorithm to search for the global optimal clustering centers is designed to accurately cluster the camouflage colors, and (3) the super-pixel algorithm which works efficiently to eliminate the noise and texture boundaries of images is used to smooth the image after color clustering. Compared to four typical algorithms, the proposed algorithm generates camouflage patterns that improve the color similarity (CSIM) by 8.3% to 38.2%, the texture similarity (TSIM) by 9.9% to 35.1%, and the edge contour similarity (ESIM) by 6.9% to 37.5%.

Protecting the personal safety of on-site workers is an important task in enterprise production. In order to achieve widespread deployment to edge computing terminals, a lightweight object detection algorithm based on YOLOv5 is used to implement the personal safety detection task for workers. To achieve a lightweight task, PConv is utilized as the convolutional layer to decrease computational complexity, while Bi-Level Routing Attention is incorporated to enhance model accuracy. Furthermore, four detection heads are employed to improve object recognition capabilities. After experimentation, the precision can be improved by 3.4% compared with the baseline model, the parameters are reduced by 1.91MB, and the model size is decreased by 3.2MB.

This study pioneers a comprehensive framework for momentum prediction, incorporating a multifaceted methodology that integrates Markov chains, Analytic Hierarchy Process (AHP)-TOPSIS, and Principal Component Analysis (PCA). This novel amalgamation enhances the robustness and depth of the predictive analysis. The use of AHP-TOPSIS augments decision-making processes by prioritizing critical factors influencing momentum, while PCA contributes to data dimensionality reduction, ensuring a focused and efficient analysis. This holistic approach not only advances the methodological landscape of sports momentum prediction but also sets a precedent for future interdisciplinary research in predictive analytics.

Discuss the objective problems and limitations found in the case of strong specificity, high discrimination requirements and regional significance. The use of the nonlinear module to increase the output of the front layer. Get a form dataset like CK and mark the corresponding number to each category for easy training. Obtain low-scale sensitive communication features to identify the sample form and conduct a practical and systematic test on the effectiveness of the algorithm. This process has a considerable long-term significance for the continued development of face recognition.

Forecasting traffic flow is a vital component in driving the advancement of modern cities and establishing intelligent transportation systems. The existing graph modeling methods have two limitations: the static spatial modeling cannot capture the fine-grained relationship between nodes, and the long-term prediction ability is poor. Aiming at these two limitations, an adaptive graph convolutional recurrent network (Ada-GCRN) is proposed. The main core is as follows: the adaptive graph structure learning component combines GCN to improve the extraction of spatial features, and the Transformer-like layer is introduced to capture global time dependence to improve the ability of long-term prediction.

Multiplayer Online Battle Arena (MOBA) games currently dominate the esports landscape, offering a concrete and vivid embodiment for team comparisons, where accurately predicting the winning team is both important and challenging. Recent studies often rely on the static features of team members to evaluate a team’s capabilities and predict their likelihood of winning. However, such approaches fail to consider the critical game-contextual influences of interactions with teammates and opponents on individuals. During a match, an individual’s performance is consistently influenced by teammates and opponents, leading to variations in individual features across different games. Moreover, the measurement of team capabilities also needs improvement, as the single scalar rating employed is too simplistic for sophisticated MOBA games. To this end, we propose a novel Context-Enhanced Match Outcome Prediction model (CEMOP), which incorporates the effect of both teammates and opponents for a better measurement of team capabilities. Specifically, we designed a match context encoder for MOBA games to evaluates the impact of teammates and opponents on individuals separately and embeds the information into the individual representations. After that the single team member’s representations which contain the influence of teammates and opponents are aggregated based on their importance to create a comprehensive representation of the team capabilities, enables CEMOP to adaptively adjust team evaluations. Experimental results from three real-world MOBA games clearly verify the effectiveness of CEMOP compared with several state-of-the-art methods.

This article addresses the challenges of data privacy protection and information security in the context of standard digitization, systematically examining the current status and practical applications of relevant algorithmic technologies. It elucidates the essence and characteristics of standard digitization, analyzes the necessity and urgency of strengthening data privacy protection and information security. The article reviews the current situation of data privacy protection and information security both domestically and internationally, highlighting the heightened requirements imposed by standard digitization on both fronts. It focuses on exploring advanced privacy protection and security algorithmic technologies such as homomorphic encryption, differential privacy, federated learning, cryptographic algorithms, artificial intelligence, and evaluates their performance through experiments. Finally, the article analyzes the application effects of algorithmic technologies in real scenarios, using a case study from the electricity industry, confirming their effectiveness and practicality. This research holds significant reference value for promoting technological innovation and practical applications in data privacy protection and information security under the conditions of digital economy.

Object detection in urban traffic scenarios is a great challenge in computer vision due of the factors like occlusion, illumination, small target size and complex background. To enhance the accuracy of object detection in urban traffic scenarios, several innovations is applied to original YOLOv8. First, a novel C2f module embedded with deformable convolution network (DCN) is proposed to increase model's receptive field and therefore strengthen the model's robustness. Then, channel priority convolution attention (CPCA) mechanism is introduced to extract important features and thus enhance the regression and localization capability of the model. Furthermore, the loss function CIoU is replaced by ECIoU to improve detection frame positioning accuracy and convergence speed. From the experimental results on the VisDrone2019 dataset, the improved YOLOv8 model has obtained better performance than original YOLO models. This study provides a theoretical basis for more accurate object detection in complex urban traffic scenarios.

Segmentation of colorectal polyps is essential for colorectal cancer prevention. The accuracy of segmentation is affected by the diversity of shapes, sizes, and textures of colorectal polyps, as well as unclear polyp boundaries. Therefore, we propose a joint HarDNetV2 and attention mechanism for polyp image segmentation network. The encoder uses HarDNetV2 to extract backbone features to improve inference speed and computational efficiency. The decoder includes an attention gate module (AG) and a convolutional attention module (CAM), in which the AG module fuses the extracted features of different sizes, the CAM module groups features into pixels and uses channel and spatial attention to suppress background information, which enhances the recognition of polyp regions and improves segmentation performance. Experimental results show that the proposed method achieves 92.17% and 88.16% performance for mDice and mIOU on the Kvasir-SEG dataset, respectively.

Combustible gases exhibit high absorption capabilities in the mid-infrared spectrum, making mid-infrared imagery a promising trend for visualizing future combustible gas leakage. However, due to the inconspicuous features of combustible gases in mid-infrared images and their susceptibility to interference from other moving objects, the detection error rate remains high. To address this, a combustible gas leakage detection algorithm that integrates foreground region detection with semantic segmentation has been devised. Initially, average background is computed using confidence intervals and recursive methods, coupled with cumulative mean and variance subimages for foreground denoising. Subsequently, adaptive threshold segmentation based on grayscale histograms is employed to identify all moving foreground regions. Furthermore, a multimodal fusion semantic segmentation network, GasLeakNet, is introduced to eliminate non-gas foreground targets. Ultimately, the proposed method achieves gas region extraction and visualization through the fusion of foreground region extraction and semantic segmentation results, utilizing adaptive threshold segmentation. This approach maximizes the strengths of foreground region detection and semantic segmentation algorithms, yielding superior results in combustible gas leakage foreground detection. Experimental results demonstrate the method’s enhanced accuracy in detecting combustible gas leakage under various conditions compared to existing foreground detection algorithms.

With the rapid development of online learning platforms, learners have more opportunities to choose courses. However, many mainstream education platforms currently lack personalized recommendation process, and due to the diversity of course quality and content, many students find it difficult to choose more suitable courses for themselves. Therefore, personalized course recommendation plays a core role in online education platforms. The existing art education curriculum has a variety of attributes, and these attributes also include a variety of modal forms (such as course video, course pictures and course introduction); Learners themselves also have a variety of attributes, their age, educational background have an impact on the recommendation results; There is also information exchange between learners and the course, such as rating on the course, which can reflect the degree of learners' interest in the course and connect them to each other. Therefore, this paper constructs a knowledge graph, which regards learners and courses as entities and the multi-modal form of courses as course attributes. Taking learners' personal information as learner attributes and different information exchanges between learners and courses as relationships .We use the applicability of graph networks to graph structures to capture learners' preferences for different modal attributes, so as to make more accurate personalized recommendations. In this paper, MOOCCube is used as the main data set, and experiments show that the proposed algorithm is better than the baseline.

In the particular environment conditions, such as haze and smoke, the quality of aerospace remote sensing images obtained can be reduced to varying degrees, affecting their subsequent usage. Currently, image dehazing algorithms are primarily categorized into three major types: image enhancement-based dehazing, image restoration-based dehazing, and deep learning-based algorithms. This article analyzes the dehazing principles, effects, advantages, and disadvantages of these three major types of remote sensing image dehazing algorithms, and also looks ahead to the future development directions of this research field.

Production safety has always been a paramount concern for coal preparation plants. With the widespread availability of video capture equipments, using monitoring videos to enhance the warning capabilities and detection functionalities for equipment malfunctions has become a priority. This paper will focus on the problem of the conveyor belt deviation detection which occurs frequently in the coal preparation plants. Nowadays, most conveyor belt edge detection algorithms can only detect straight edges and cannot adapt to edge shapes. In addition, some algorithms cannot be applied in scenarios with external influences such as strong changes in illumination, excessive image exposure, or water accumulation. In response to these shortcomings, this paper proposes an algorithm with enhanced robustness that can adaptively detect conveyor belt edges. Experiments show that this algorithm has good illumination equalization effect, edge detection effect and accuracy, deviation detection accuracy, and works well in various situations. At the same time, this algorithm has already been successfully implemented for use in coal mining production.

To improve recommendation systems' accuracy and customization while preserving user privacy, this study introduces a federated learning-based dynamic recommendation model, D2PDRF. Utilizing an attention mechanism, the model adjusts the blend of long-term and short-term interests, enhancing system adaptability. By integrating differential privacy with federated learning, user privacy is safeguarded. Tests on public datasets show the model enhances recommendation precision and personalization, ensuring data confidentiality.

With the acceleration of social development and industrialization, the problems of traffic efficiency due to road congestion, traffic safety due to unregulated driving, and environmental pollution due to increased carbon emissions are all becoming more and more serious. With the development of 5th generation mobile communication technology (5G), a cellular-vehicle- to-everything (C-V2X) technology can be used, C-V2X) technology can provide vehicles with long-distance and wide-range wireless network communication. With the development of machine learning and high-precision mapping technologies, L4 (SAE International's definition of autonomous driving technology classification) autonomous driving technology allows vehicles to be driverless in certain environments and under certain conditions. both C-V2X and autonomous driving technologies offer fruitful solutions to reduce road congestion, erratic driving, and carbon emissions from vehicle emissions. C-V2X and autonomous driving technologies both provide fruitful solutions for reducing road congestion, irregular driving and reducing carbon emissions from vehicle exhausts, and also bring new ideas for improving transportation efficiency. In this paper, we propose an implementation of an intelligent transportation system based on 5G, C-V2X and L4-level autonomous driving technology. In this intelligent transportation system, the discussion focuses on the Vehicle-Vehicle-Circuit Cloud Collaboration (VVC) system applicable to the public transportation sector and the value played by the VVC system, which not only facilitates the commercialization of 5G and automated driving, but also establishes a new set of business models suitable for the current transportation industry, which is of great significance for improving transportation efficiency and reducing carbon emissions.

Two-dimensional textbook cannot show the three-dimensional effect of solid geometry. For many middle school students who lack three-dimensional space thinking, a teaching assistance system that can vividly display three-dimensional geometry and interact with it is essential. In recent years, the introduction of augmented reality technology into education and teaching has attracted much attention from the industry. However, the existing augmented reality teaching assistance system either does not support interaction, resulting in poor teaching effect; Or they need to wear VR devices to achieve interaction, resulting in excessive physical burden on students and affecting the learning experience; Or an app-based approach, which requires downloading before it can be used, is difficult to widely serve students. In response to the above problems, this paper completed the system design by combining web technology and augmented reality technology, and realized real-time interaction between human hands and virtual three-dimensional geometry, solving the problem of weak interactivity of traditional augmented reality teaching assistance systems. Finally, through a questionnaire survey, the user experience of this system is compared with that of traditional textbooks, and the results show that this system has a better learning experience.

With the advancement and maturity of computer vision technology, its application in the field of education has become a new trend. This paper presents a research study on the improvement of YOLOv7 for classroom student object detection, which holds significant implications for integrating computer vision technology into the educational domain. During the detection of classroom student targets using YOLOv7, challenges such as dense distribution, severe occlusion, and small target sizes were identified. These challenges led to issues like missed detections and false positives[1]. Addressing these issues, this paper introduces a repulsion loss function, building upon the original loss function of YOLOv7, to enhance the regression of bounding boxes. Training on a custom classroom student dataset revealed an improvement in the detection performance of YOLOv7 after incorporating the modified loss function. In comparison to the original model, the enhanced YOLOv7 achieved an average precision of 78.8%, representing a 0.8% improvement. The results indicate that the addition of the repulsion loss function enhances the accuracy of YOLOv7 in detecting student targets in real classroom environments.

Classical Chinese poetry employs poetic imagery extensively to convey feelings. To improve the effect of poetry sentiment analysis, this paper introduces imagery features into the model and proposes an improved BiLSTM poetry sentiment analysis model. The experimental results demonstrate that the improved BiLSTM model incorporating imagery features achieves better results than the traditional BiLSTM model in the task of poetry sentiment analysis.

Skyline query is a prevalent multi-objective optimization problem designed to identify points in a given dataset that are not dominated by any other points. Although it traditionally emphasizes the analysis of individual points, real-life scenarios often demand consideration of combined queries, given the necessity for optimal combinations. In addressing this, we explore the Skyline query algorithm with a focus on grouping, presenting an efficient G-Skyline query algorithm. Our approach involves a streamlined batch processing algorithm that groups all points to compute the skyline points for the initial l layers. Experimental results demonstrate that our algorithms outperform previous work in terms of speed.

Story generation is an important part of natural language processing, text generation tasks have received extensive attention from scholars in recent years. Existing methods are difficult to generate complete, logically fluent texts with low repetition rate based on a given outline vocabulary. To solve this problem, based on the dataset generated from outline conditional stories, we use the latest GLM model and instruct-tuning to generate Chinese fables and compare it with other big model fine-tuning with different parameter settings, and the experimental results show that the fables generated by Instruct-tuning have stronger logic and lower repetition rate. The models we used showed an increase of 1.5%, 2.43%, 1.51%, 3.25%, and 2.09% over the baseline model on the Blue-2, distinct-3, distinct-2, distinct-3, and distinct-4 metrics.

In order to quickly adapt to the changes of tasks and the innovation of information technology, it is required that an information system needs to be continuously optimized and improved during its construction and service phases. Continuous evolution has become the basic characteristics and requirements of an information system. How to manage the requirements in the continuous evolution process of the system, to improve the efficiency and quality of iterative evolution, has become a key problem that needs to be solved urgently. Therefore, this paper proposes a refined requirement management method for Continuously Evolving Information Systems (CEIS). First, via comprehensive analysis of evolution process of CEIS, a requirement life cycle process which contains 4 stages, 17 activities and 14 data, is proposed; Second, the 14 requirement-related data are further classified and refined to form a requirement meta-model of CEIS for defining and describing their relationships in a standard way; Third, based on the requirement life cycle process and metamodel, an refined requirement management process is designed, and 21 main activities of requirement management are summarized; Finally, a requirement management tool is developed based on the proposed method, and the main management interface is displayed with an application to illustrate the effectiveness of this method.

In response to the challenge that traditional camouflage design methods struggle to evade detection by modern unmanned aerial reconnaissance, we propose a camouflage pattern generation adversarial network model using color-semantic constraints. A reference image generation model using color-semantic constraints has been established, which through sentence encoding model, generates reference images with fundamental texture and color features. This is achieved by incorporating adversarial loss, texture loss, and pixel-level loss. We design a color standardization processing strategy based on the SimCLR framework. This model generates semantic camouflage images in batches by designing data augmentation strategies, positive-negative sample similarity measurement strategies, and sample structural similarity algorithms, regard to reference images. Qualitative and quantitative experimental results demonstrate that our proposed method exhibits strong camouflage performance in different environmental settings.

Burst Buffer was proposed to address the growing performance gap between computation and I/O in high-performance computing systems (HPC). However, the introduction of Burst Buffer has brought new issues in system resource management and job scheduling. Resource management systems on HPC platforms lack efficient management of multiple types of resources, and the allocation of computing resources and Burst Buffer resources is independent of each other. This can lead to underutilization of system resources and job blocking. To address these issues, this paper proposes GABB, a plan-based job scheduling strategy for shared Burst Buffer, and utilizes genetic algorithms for optimization. GABB employs a plan-based job scheduling strategy to uniformly schedule all jobs in the waiting queue to generate an execution plan. Furthermore, GABB comprehensively considers changes in system resources and the job's demand for computing and Burst Buffer resources during the job scheduling process. Finally, GABB utilizes an improved genetic algorithm to optimize the job scheduling scheme. We conducted experimental simulations of a shared Burst Buffer system and implemented a plan-based job scheduling algorithm. The experimental results indicate that BBGA significantly reduces the mean waiting time of jobs by over 20% compared to the Shortest Job First (SJF) algorithm, and reduces the mean bounded slowdown of jobs by over 25%.

LSM-tree is favored in write-intensive scenarios due to its sequential write feature, but it grapples with read and write amplification issues. A common solution is to store keys and values separately, with the LSM-tree holding only the key and value address, while the value is kept separately. However, this approach presents challenges in range query and garbage collection, necessitating multiple random reads and pointless data rewriting. To address these issues, we introduce TreeKV, a system that uses a specialized B+ tree for value storage, grouping values with adjacent keys together. This reduces random reads during range queries and eliminates unnecessary data rewriting during garbage collection. We also delve into crash recovery and minimal-blocking garbage collection implementation in TreeKV. Our comparison with existing key-value separated stores shows that TreeKV, while maintaining high read-write performance, enhances range query and garbage collection performance.

This paper explores using the AIGC tool ChatGPT to enhance interdisciplinary teaching in colleges and universities. Interdisciplinary courses, crucial for nurturing well-rounded talents, face challenges such as teachers' limited interdisciplinary knowledge, outdated teaching resources, and students' difficulties grasping complex concepts. ChatGPT offers extensive knowledge, instant response, and personalized learning to address these issues, supporting teachers with interdisciplinary teaching resources, instructional design, and assessment tools while providing students with customized learning materials. The proposed Interdisciplinary Practice Teaching Solution, comprising four core modules, facilitates the organic integration of interdisciplinary learning. A practical project, "Portfolio Analysis of Stock Funds," is used as a case study to demonstrate ChatGPT-aided interdisciplinary teaching. It showcases its effectiveness in cultivating students' abilities to deconstruct projects, clarify concepts, provide targeted hints, and critically validate and integrate projects. Experimental data confirm the framework's excellent performance across various indicators.

In the pursuit of advanced human-machine interactions, the ability to detect emotions in textual data emerges as a crucial element for imbuing machines with empathetic communication capabilities. This paper proposes a theoretical framework for the Convolution Selective State Space Model (ConvSSM), a deep learning model designed to discern and classify the emotions conveyed through text. Unlike conventional analysis models, the ConvSSM is designed to accommodate a wide array of emotional expressions, thereby capturing the complexity inherent in textual emotional states. Experiments show that our model has better performance.

Deep learning has found extensive applications in the domain of autonomous driving. However, in complex road environments, diverse obstacles such as irregularly shaped objects, children's toys, animals, and other unconventional entities pose significant challenges. Convolutional Neural Network (CNN)-based road detectors struggle to satisfy real-time demands owing to the complexities associated with accommodating multi-scale and intricate backgrounds. In this paper, for the road obstacle detection problem in the field of autonomous driving, we propose a YOLOv8-based detection method, ROD-YOLO (Road Obstacle Detection), which has a better multi-scale adaptability, and the model is used to segment the obstacles on the road. Compared to the original network, ROD-YOLO adds a detection header, and this paper proposes a method to add Transfomer with GAM attention mechanism to the C2f module. In order to make the model better adapt to multi-scale obstacles, we add a new small-scale segmentation header and a special feature fusion part. Specifically the new GlobalCSP C2FGAM module is proposed with the C2STR module that incorporates the Transfomer idea to obtain faster segmentation speed and better accuracy for different obstacles, and the algorithm performs well in real-time object segmentation tasks and is able to maintain a high level of accuracy. It improves the mAP by 1.9% compared to the original network YOLOv8, which significantly improves the segmentation of small object samples. The research results in this paper are of great significance for improving the safety and efficiency of self-driving vehicles.

Environmental monitoring has become a very important link in environmental management and protection. In this paper, an Android-based environmental monitoring information management system has been designed, organically integrating the key processes and resource conditions of environmental monitoring into the system for effective management and control. Furthermore, using mobile design, it has been more efficient, flexible, and convenient. The system is designed in the MVC pattern and divided into five levels: basic layer, data layer, support layer, application layer, and presentation layer, while including five functional modules as follows: business process management, resource management, data management, sample management and system management, achieving automation in environmental monitoring processes and dynamic resource management. The implementation of this system enables environmental monitoring laboratories to enhance the management level, greatly reduce management costs, ensure the accuracy and safety of experimental data, and therefore have great promotion value.

It is not a valid asseration that everyone can paint as the difference of people's drawing ability. Thus, there is a common problem of 'bad painting' . Therefore, how to automatically optimize a large number of real and low-quality hand-drawn sketches produced in actual situations such as people 'can't draw' and 'can't draw well', so as to help most people reduce the pressure of painting, and achieve their own painting expectations when doodling, and more importantly, enhance the discriminability, interpretability and applicability of such hand-drawn sketches in vector style, is a critical problem to be solved in the field of freehand sketch processing. In this paper, we intend to address the issue of optimizing low-quality vector sketches based on diffusion model. Concretely, we first propose a criterion of information entropy (IE) to define sketch quality and use it as a preprocessing method to preliminarily screen sketch data, a latent structure which based on diffusion model is then utilized to optimize the low-quality sketch. Extensive experiments demonstrate the effectiveness of the proposed method, furthermore, the optimized sketch can be used for downstream tasks such as image recognition and understanding.

In today's era of advancement, the significance of digital technology and data resources in government governance has escalated. This article introduces the CNTM model, which stands as a novel topic model incorporating convolutional neural network (CNN) and long short-term memory network (LSTM). It can effectively model topic discovery and topic evolution in information networks. First, the CNTM model achieves in-depth representation of topics in information networks by integrating the advantages of CNN and LSTM and combining text content and network structure information. Secondly, the CNTM model can use topic model learning to predict and analyze the process and results of information dissemination, helping government agencies understand the evolution trends of social public opinion and public topics, and respond and intervene in a timely manner. To maximize the impact of the CNTM model in digital governance, this paper enhances its structure and algorithm, aiming to enhance its capability in processing extensive network data. It also optimizes the model parameters to improve the accuracy of topic discovery and information dissemination prediction. To sum up, the CNTM model, as a new topic model based on CNN and LSTM, demonstrates a powerful mechanism and optimization strategy in digital governance, provides effective information tools for government agencies, and promotes the intelligence and precise development of digital governance.

Deep learning has achieved excellent results in semantic segmentation in computer vision, but this requires a large amount of data and computational power. Therefore, Class Incremental Semantic Segmentation (CISS) is becoming one of the current research hotspots, which involves gradually learning new classes in the incremental step to update the model. However, the existence of catastrophic forgetting can cause the model to forget the previously learned classes, especially when background class pixels are offset, which further exacerbates this problem in CISS. To address this issue, this article proposes a new approach that combines knowledge distillation and latent class comparison loss to alleviate catastrophic forgetting. The effectiveness of our method has been demonstrated through extensive experiments in existing CISS tasks and recently proposed challenging tasks.

In today's information society, the importance of collecting data through multiple channels continues to increase. Among them, the use of drones to collect information on large-area forests is in line with the modern green development concept. In actual drone collection tasks, it is crucial to ensure the accuracy of collected data and the safe flight of the drone. In order to deal with these challenges, a multi-sensor fusion algorithm based on Bayesian estimation is proposed to overcome the problems of insufficient data accuracy of a single sensor and data duplication and noise caused by simple fusion of multiple sensors. At the same time, considering the actual need for path planning without a global map when flying over a large area in the woods, a local grid mapping strategy based on sliding windows is proposed, that is, a moving map is established with the drone body as the map center to ensure the maintenance of a fixed map size. This strategy effectively reduces the memory usage problem and is consistent with the actual scenario of large-scale outdoor flight.

Micro service for varieties of micro application is a crucial approach to achieve the success of smart campus construction. The research is from the perspective of micro service on smart campus, and taking a classic graphic game of "Tank Battle" development as an example. This study puts forward a feasible construction solution of game program based on Spring cloud for micro service for the purpose of recreation and entertainment, and devotes to completing rapid development by Scrum model of Agile method. A feasible and scientific Agile design process is described which consists of development tool of QT selection, micro service architecture design, game scenario description, major algorithm on behaviors of game elements, user story division on requirement acquisition and visual game map implementation. Key game elements are implemented effectively for the first Sprint of game version 1.0, which runs stability and positively for players. Application result shows that game UI is friendly and convenient to operate, which is capable of attracting students' interests on leisure time on campus.

This paper for enterprises, scientific research institutes and other kinds of organization to implement the product life cycle management needing, according to the difficulties and obstacles existing in the practice of product quality control, put forward the whole life cycle management of product quality information system construction ideas and implementation points, aims to organize the implementation of product life cycle information management put forward feasible solutions.

Medical Named Entity Recognition (NER) plays a crucial role in enhancing the efficiency of clinical work. Currently, Chinese Named Entity Recognition methods based on deep learning models have shown significant results in this area. However, due to the differences between Chinese and English, many methods that perform well on English datasets cannot be directly transferred to the Chinese context. Additionally, identifying medical entities also poses difficulties due to the comparative scarcity of specialized medical expertise. Furthermore, the prevalent models, which are typically tailored for recognizing non-nested entities, fall short of accurately identifying nested entities within Chinese medical texts. To address these issues, we propose a Chinese medical named entity recognition model based on feature fusion and multi-head biaffine transformations. By utilizing multi-head biaffine transformations to construct span matrices and applying convolutional networks on each channel to fully model adjacent span information, we solve the problem of nested entity recognition. Additionally, we significantly enhance the accuracy of medical entity recognition by introducing Chinese medical word vectors. Finally, our research involved testing on both a nested Chinese medical entity dataset (CBLUE-CMeEE) and a non-nested medical entity dataset (CCKS2019-Yidu-S4K). The experimental results show that our proposed model improves across all metrics, becoming the new state-of-the-art (SOTA) model.

With the arrival of the digital era, the traditional human resource management model has been unable to meet the development needs of enterprises, human resource management informatization construction is particularly important to improve the effectiveness of enterprise management. This paper analyzes the current situation of human resource management in small and medium-sized enterprises(SMEs) and the necessity of informatization construction, puts forward the countermeasures of human resource management informatization construction in SMEs, which lays the theoretical foundation for the construction of human resource management information system in SMEs with the goal of digitalized information management, and finally builds a demonstration system under the guidance of countermeasures.

To solve the problem of low load recognition accuracy caused by high load feature dimension in non-invasive load recognition, a non-invasive load recognition method combining random forest (RF) and K-nearest neighbor (KNN) was proposed. Firstly, feature extraction is carried out on the current data collected by the equipment during stable operation. Then, RF is used to score the feature importance of the load characteristics. Finally, the data with high load feature scores are selected and input into the KNN model to complete load identification and classification. The experimental results show that the random forest (RF) feature selection for high-dimensional load features can effectively improve the load identification accuracy.

In Automatic Speech Recognition (ASR) tasks with end-to-end speech recognition models based on deep learning, some critical information may only exist in the local region of the input sequence. Therefore, it is not only necessary for the model to capture global speech features for interaction, but also need to extract and model the local features in a reasonable way. Meanwhile, due to the high complexity of Transformer model computation, it will face some challenges in practical applications. To address the above problems, this paper proposes a PREFormer model based on Transformer, which incorporates pointwise convolution into the feedforward layer, and at the same time adds residual connection between adjacent Attention modules for feature fusion. Finally, it introduces softmax with temperature to optimize the model's handling of probability distribution during training. The experimental data show that the model in this paper reduces the Character Error Rate (CER) by 4.14% on the Aishell-1 data set with better robustness compared to the conventional ASR model.

Cognitive diagnosis is a critical focus for smart education systems, with the goal of evaluating students’ cognitive states and predicting scores based on their responses. Recent studies have primarily concentrated on leveraging neural networks to forge student-exercise interactions models. However, further exploration is needed to identify effective strategies for distinguishing correct and incorrect student responses. To address this challenge, we propose a Dual-channel Cognitive Diagnosis (DCD) framework that enhances features to determine students’ cognitive states by analyzing correct and incorrect responses. Initially, we utilize two types of response data to create adjacency matrices for heterogeneous graphs, distinguishing the characteristics of accurate and erroneous student responses. Subsequently, the students’ cognitive states are inferred by modeling two channels: the correct response channel and the incorrect response channel. Ultimately, the outcomes of experiments conducted across diverse authentic datasets validate the efficacy of our proposed framework.

As the advancement of computer technique and artificial intelligence technology, education based on computer technology has become a hot topic, and the wave of educational informatization is rising. Among them, adaptive learning has received widespread attention as a type of educational informatization technology. However, the current adaptive learning feature models use different semantics to describe each other, and the models lack meta model level explanations, resulting in a lack of further operability of the models and difficulty in integrating with each other. In order to solve these problems, our paper proposes a fusion and verification method for learning feature adaptive models based on meta models. Above all, an adaptive learning feature meta model is proposed, which defines the construction rules of the adaptive learning feature model at meta model level, so that different models can be fused using the same set of semantic rules. Then, at the meta model level, transform the learning model into a requirement model. Finally, using formal methods, validate the fused feature model, which can make sure of the semantic consistency between the previous and subsequent models. This article demonstrates the entire process through a model fusion experiment. The experimental results indicate that this method is correct and feasible in promoting feature model fusion without adaptive learning.

In the field of industrial control, a large number of display and control equipment are required to provide display interface and human-computer interaction interface to realize the implementation and monitoring of the entire process by technicians. The robustness design of the display link is very important, this paper carries out the scheme design of the industrial control display and control equipment, sorts out several factors that restrict the robustness of the display link, and finally analyzes and troubleshoots the display problems encountered in the actual project.

With the processes of economic globalization and social technologization, the Internet of Things, cloud computing, and artificial intelligence are changing people's lifestyles. The concept of "smart hotels" has also rapidly emerged, providing an effective way for the transformation of the traditional hotel industry. Through research, it was found that the construction status of smart hotels is uneven, lacking systematicity and completeness. This paper, based on the SAPAD theoretical model and the AHP hierarchical analysis method, analyzes the key behaviors of users during the hotel check-in process from the perspective of service design, explores potential customer needs, reconstructs the functional modules of hotel accommodation services, and aims to enhance the hotel user's check-in experience, improve the smart hotel service system, and increase hotel user loyalty.

Temporal expression extraction is to identify temporal expressions from raw text. Predominant methods rely on handcrafted rules or high-quality training datasets, and currently few works utilize large language models in this task. This paper demonstrated the ability of large language models to understand temporal information by building few-shot temporal expression extraction system that does not require hand-crafted rules or training. To reduce model output length, the system used phrase-only output schema instead of outputting entire sequence. The model also alleviated hallucination effect by prompting the model to predict expression type and using the model itself to filter model outputs. Experimental results indicated that the model achieved decent performance without training or rules.

Many multi-source transfer learning methods use the latent attributes between source domains to mitigate interference. However, these strategies have some shortcomings. Strategies that only use commonality may lose valuable latent information specific to a single source domain when extracting knowledge. Strategies that focus on specificity are limited by the differences between source domains, so they cannot effectively integrate different knowledge. For strategies that include both attributes, they only explore the knowledge scattered in the original source domains, and do not attempt to explore additional knowledge, which can fill the learning gap and promote knowledge fusion. Based on the above problems, this paper proposes a multi-source transfer learning method that can deeply mine the commonality and characteristics of multiple source domains and fully integrate all the latent information by reconstructing the multisource latent feature space and combining high-level concept learning. Finally, we conduct a large number of experiments to verify the effectiveness and superiority of the algorithm.

The grading of tobacco leaves is a pivotal step in the tobacco production pipeline. Traditionally, this grading has been performed manually by seasoned experts, a practice both time-intensive and reliant on subjective assessments. In recent years, advancements have been made towards automating this procedure through the analysis of visible light images of the leaves. However, due to the high visual similarity among visible light images at different grades, a single visible light image is insufficient for achieving accurate grading. It is known that chemical composition serves as a pivotal metric for evaluating tobacco leaf quality, where the near-infrared spectral image of tobacco leaves probably harbors valuable information pertinent to grading. Inspired by this, we propose an end-to-end multispectral feature enhancement network for automated tobacco leaf grading. In addition to utilizing common visible light images, the network integrates a principal component from the near-infrared spectral image to encapsulate chemical information. Detailed experimental results demonstrate that our method achieves a high grading accuracy of 91.94% with the inclusion of near-infrared spectral images, surpassing the performance of existing methods reliant only on visible light images. The network showcases high accuracy while maintaining swift inference speed, offering innovative insights for the future design of automated tobacco leaf grading systems.

In the process of research and development of automotive products, the man-machine layout design of cab comfort has become the focus of development. At present, to judge whet that man-machine arrangement of the cab of the commercial vehicle meets the comfort requirement, a sample vehicle is usually manufacture first, Evaluated by professionals, this development process is not only cumbersome, high error rate and low efficiency, but also consumes huge human and financial resources. In this paper, the man-machine bench test of commercial vehicle is designed and carried out. The corresponding comfort evaluation data of each percentile driver under different man-machine arrangements are collected. Secondly, a RBF neural network model is built to predict the subjective evaluation of man-machine layout. The training set and test set of the model use the data from the bench test, and the accuracy of the model is evaluated by objective indicators. Finally, the secondary development of the commonly used design software CATIA is carried out, and the program is written to automatically read the size information of the three-dimensional model. One key evaluation is realized by embedding RBF neural network. The example shows that the software of RBF neural network and CATIA secondary development can greatly improve the efficiency of designers.

This study is dedicated to the development and evaluation of a large language model specifically optimized for the biomedical domain, using the sophisticated Llama2 architecture as its foundation. By delving into the vast corpus of biomedical literature, which includes medical research papers, news reports, drug labels, and case studies, we have curated a comprehensive training dataset comprising 10 billion tokens aimed at addressing the deficiencies of existing language models in specialized medical knowledge. During the model training phase, we initially conducted pre-training using data from the biomedical domain, followed by supervised training of a biomedical question-answering model utilizing LoRA technology. In the evaluation phase, the model's performance underwent rigorous testing with challenging datasets, including the Chinese National Medical Licensing Examination, the United States Pharmacist Licensing Examination, and the PubMedQA question-answering dataset. These tests served as a robust indicator of the model's practical applicability in real-world medical professional scenarios. Experimental results suggest that our model performs exceptionally well in various specific biomedical tasks and surpasses the well-known ChatGPT-3.5 model in key performance metrics, affirming the effectiveness of domain-optimized large language models. Despite the exemplary performance, the model still lags behind the latest ChatGPT-4, which indicates potential for improvement in understanding depth, reasoning capabilities, and knowledge updating. Overall, this research provides a potent tool for natural language processing tasks in the biomedical field and lays the groundwork for the future development of more specialized and precise large language models.

Planning management is the most important aspect of management activities in large manufacturing enterprises. In addition to their large scale, large manufacturing enterprises are also reflected in the complexity of their manufacturing environment, such as multiple workshops with different specialties, multiple production projects, complex market environments, and so on. How to use a planning management model to effectively undertake project plans upwards, penetrate into end plans downwards, and carry out efficient management has become an important research topic for modern enterprises. This article proposes a new multi-level planning management structure and deploys a new priority algorithm at the workshop planning layer. This algorithm focuses on factors such as device loading and task preemption in the earliest deadline priority scheduling algorithm EDF, and has improvements in reducing task loss and aligning with actual production. In addition, innovation focuses on the study of planning, assessment, and management models in the field of scheduling research. The article proposes a plan performance model that can accurately evaluate project teams of different scales, resulting in a 14.74% increase in the plan execution rate of a certain enterprise.

In the process of recognizing dangerous assembly behavior in machine tools, there are challenges such as the difficulty of behavior interval calibration and the limited feature discrimination ability of existing recognition models. To address these challenges, a cascade network method for dangerous assembly behavior recognition is proposed. First, this paper proposes the behavior category hypothesis module, dual-channel structures with different sampling frame rates were used to learn human action features with spatio-temporal semantic information, providing supervised behavior category labels for the fine-grained classification in the second stage; Secondly, this paper constructs the behavior information detection module and incorporates a target and attitude detection phase based on the results obtained from the first stage. This is achieved by capturing key behavioral information features that influence the discrimination of behaviors in video through FPN and PAN structures; Finally, the stage decision algorithm of the behavior definition module realizes the multimodal feature fusion of human action and behavior information, and then the rapid decision of dangerous assembly behavior. The CART decision tree algorithm is used to construct the feature weighting structure, which improves the accuracy of dangerous assembly behavior recognition. Experiments show that the proposed method can reach 93.6% accuracy with good accuracy and robustness.

The domain of palmprint recognition, characterized by its convenience, low privacy sensitivity, and rich feature sets, has garnered increasing research interest. Moreover, Vision Transformers (ViTs) have emerged as a promising alternative to conventional convolutional neural networks, showcasing state-of-the-art or competitive performance in various downstream image processing tasks. This paradigm shift highlights the transformative potential of ViTs in areas traditionally dominated by other neural network architectures. Despite the promising avenues opened up by ViTs, their application within the palmprint recognition field remains largely unexplored, marking a significant research gap. This paper diverges from the general analysis of ViTs to focus explicitly on their application in the domain of palmprint recognition. By undertaking a comprehensive evaluation, it establishes foundational work in assessing the effectiveness of ViT-based methodologies for palmprint recognition tasks. Concluding with a forward-looking perspective, the paper suggests future directions for research that could expanding the capabilities of ViTs in specialized areas of palmprint recognition.

Deep learning has exhibited exceptional effectiveness in predicting the rupture risk of intracranial aneurysms(IA). Our study aimed to provide a valuable tool for extracting Deep Learning features and predicting the rupture risk of intracranial aneurysms. To fulfill this objective, we meticulously developed a deep learning feature extraction framework to extract deep learning features from CTA images and precisely forecast the rupture risk of intracranial aneurysms. The study included 327 IAs in the training dataset and 82 in the test dataset. A model using ResNet-34 demonstrated outstanding performance, achieving an accuracy of 87.65% (80.49%-94.82%), a recall of 98.33% (95.55%-98.74%), and an AUC of 84.52% (76.65%-92.40%). Our results showcase that assessing aneurysm rupture risk performs well when utilizing deep learning features extracted through our developed framework.

In order to further improve China's forest coverage, develop ecological culture, cope with adverse weather conditions, mechanized afforestation has become the only way to achieve China's ecological civilization. This study uses the concept of small and medium-sized robots to design an intelligent tree planting robot based on Robot Operating System (ROS), which is suitable for complex terrain such as sand and stone wasteland and abandoned mines. The robot has the function of independent mapping, path planning and navigation, and can automatically complete the tree planting work in the two-dimensional plane. The overall structure of this robot consists of three parts: walking mechanism, seedling storage mechanism, and actuating mechanism. It is mainly used for ecological restoration of barren land. It has a high degree of functional integration and operation automation, and realizes the planting of arbor ecological restoration forest seedlings (such as Hongsenhuai) under the "man machine" collaborative mode.

The prompt and accurate diagnosis of brain tumors is crucial for favorable patient prognoses. Advances in computer vision, particularly deep learning algorithms, have ushered in innovative methods for enhancing the diagnostic accuracy and automation of medical image analysis. You Only Look Once (YOLO) is a commonly employed technique in the field of image detection. And in this study, we used Ghost Convolution (GhostConv) enhances the YOLOv5 model by incorporating a new convolutional module to improve feature extraction, thereby enhancing model accuracy and detection speed. Through experiments, we have found that our model can achieve high accuracy and is suitable for practical medical detection.

Automatic defect detection of steel pipe weld is of great significance in industry, national defense and scientific research field. To ensure the welding quality of steel pipe, it is necessary to efficiently and accurately defect the welding defect detection. A modified attention U-Net (MAU-Net) is presented for defect detection of steel pipe weld. In the model, an attention layer is introduced as a bridge between the encoder and decoder paths, a parallel pooling attention (PPA) module is used to connect the convolution features corresponding final downsampling layer and first upsampling layer. It is verified on the steel pipe weld defect image dataset that it is effective and feasible, and is superior not only to U-Net and its variants, but also to the latest automatic and semi-automatic segmentation/annotation models of other standards.

Tumor Mutation Burden (TMB) is a critical biomarker for predicting the response to cancer im- munotherapy. Typically, TMB is usually evaluated using Whole-Exome Sequencing (WES) in clinical practice. However, this approach requires use of finite tissue specimens and time-consuming laboratory processes. To address these issues, we have developed a deep learning model called DBFormer, which can predict TMB using readily available histopathological images of lung adenocarcinoma. DBFormer is a four-tier classification pyramid structure that incorporates color convolution and Bi- Routing At- tention and consists of two primary components. The first part is the color deconvolution layer. Each patch passes through the color deconvolution layer to produces an image that combines RGB and HED image information. The second component has four stages, each with a maximum pooling layer and a DBFormer block. The maximum pooling layer reduces the image size while increasing the feature matrix’s dimension. The DBFormer block extracts and classifies the characteristics of the image. We selected 337 and 200 whole slide images (WSIs) of lung cancer from the TCGALUAD dataset as bi- nary and triple classification datasets, respectively. The experiments conducted on the aforementioned dataset demonstrate the efficacy of our model in comparison to classical deep learning models, and es- tablish its superiority over state-of-the-art methods. DBFormer achieves an area under curve (AUC)of 0.997 in the binary classification dataset and an average accuracy of 0.973 in the ternary classification dataset. Source code is available at https://github.com/yyyyxxxyyyy/DBFormer

Underwater target detection is an important method for marine life detection. However, the accuracy of target detection and recognition is affected by the problems of image occlusion, blurred water quality and complex background of underwater targets. An improved YOLOV7 underwater biological target detection method is proposed to address the above problems. Firstly, the DCN structure is improved by incorporating the residual structure into the DCN, with new jump connections and 1×1 convolutional branches, so that the network can learn the residual information between the input and the output. It can improve the expressive ability of the model, reduce the number of parameters, and enhance the adaptability to the shape of the object. Finally, COT3 is integrated, which utilises a self-attention mechanism to weight the input feature maps so that the model can focus on important contextual information around the target and improve the model's ability to handle complex scenes. The experimental results show that the improved YOLOv7 model achieves an average detection accuracy mAP value of 80.4%, which is 2.8% higher than the original algorithm, and the algorithm is better than the original YOLOv7 algorithm in terms of detection accuracy and speed, which is more suitable for the detection of underwater targets.

Intelligent Transportation Systems (ITS) have a pressing need for efficient and reliable traffic surveillance solutions. This paper for the first time proposes a surveillance system that utilizes low-cost magnetic sensors for detecting and tracking vehicles continuously along the road. The system uses multiple sensors mounted along the roadside and lane boundaries to capture the movement of vehicles. Real-time measurement data is collected by base stations and processed to produce vehicle trajectories that include position, timestamp, and speed. To address the challenge of tracking vehicles continuously on a road network using a large amount of unlabeled magnetic sensor measurements, we first define a vehicle trajectory tracking problem. We then propose a graph-based data association algorithm to track each detected vehicle, and design a related online algorithm framework respectively. We finally validate the performance via both experimental simulation and real-world road deployment. The experimental results demonstrate that the proposed solution provides a cost-effective solution to capture the driving status of vehicles and on that basis form various traffic safety and efficiency applications.

Cardiovascular diseases (CVD) are a leading cause of global mortality, as documented by the World Health Organization (WHO), resulting in a significant annual death toll. Accurate forecasting of CVD has the potential to significantly improve patient outcomes and potentially save lives. The advent of the COVID-19 pandemic has necessitated the adaptation of traditional disease prediction methodologies, underscoring the importance of a unified and precise predictive system. The integration of Artificial Intelligence (AI) in disease diagnosis and identification has been transformative, largely due to the rapid progress and reliability of deep learning technologies. This study proposes the application of Convolutional Neural Networks (CNNs) to the heart disease dataset from the University of California, Irvine (UCI) repository, employing machine learning (ML) techniques for the early detection of CVD. The analysis encompasses 14 critical attributes from the dataset. The performance of the proposed model is assessed using accuracy and confusion matrix evaluation metrics, which are employed to validate multiple prospective findings. To refine the dataset, the Isolation Forest algorithm is utilized to eliminate irrelevant features, and data standardization techniques are implemented to enhance the model's predictive accuracy. The application of deep learning methodologies has resulted in an impressive accuracy rate of 98%.

Three-dimensional (3D) reconstruction has been widely used in bone correction. However, some key problems need to be urgently solved in the current 3D reconstruction algorithms for lower extremity bones based on CT DICOM images. A method of semi-automatic 3D reconstruction and entity separation for lower extremity bones is proposed. Firstly, an improved aT-Mask (adaptiveThreshold-Mask) algorithm is researched to remove the patient couch. Then, the S-PP (Sample-Point Picking) bone-separation algorithm is proposed to well extract and separate the lower extremity bone entities. To automatically batch-repair holes on the reconstructed 3D models, the M-IM (Model-Image Matching) hole repair algorithm is implemented based on the spectral characteristics of the lower extremity bone in the DICOM images. Finally, the surface-rendering algorithm is applied to reconstruct the 3D model of the lower extremity bone. One hundred sets of DICOM images are utilized to experiment, and the fine 3D models for femur, fibula and tibia are well reconstructed and separated for each case. The 3D reconstruction of lower extremity bones can be easily established, and the entities can be separated efficiently based on the method proposed in this study. The efficiency and accuracy of deformity recognition and correction can be significantly improved by using the reconstructed 3D models.

In steel production, the quantity of surface defects is a key indicator. Traditional detection methods mainly use flash frequency to detect defects, but this method has issues such as slow speed and low accuracy. Therefore, it is of great research significance to improve the detection speed and accuracy of surface defects in steel. The main focus of this paper is to explore the task of detecting surface defects in steel, and we will first improve the YOLOv8 model to enhance its performance in detecting surface defects on steel. We will optimize the loss function calculation method, neck end, and backbone network. The specific improvements are as follows: To address issues such as dense distribution and easy overlap of annotation boxes for steel surface defect targets, we introduce deformable convolution modules into the backbone network of YOLOv8. By adding deformable convolution modules, we can flexibly handle situations where detection point receptive fields are insufficient and enhance attention towards detection targets. Secondly, replacing PANet feature pyramid with BiFPN, network allows for more effective capture of target boundary information and improves detection efficiency through multi-scale semantic feature fusion. Finally, in order to accelerate model convergence speed and improve regression accuracy when improving the YOLOv8 model, we decide to replace the CIoU regression loss function used in the original model with EIoU (Efficient Intersection over Union) loss function.

This study aims to develop a diabetes prediction platform based on Sklearn, integrating commonly used machine learning algorithms. By analyzing and modeling global diabetes data, the platform provides efficient predictions of disease probability and intuitive visualizations. The online website, built using the Flask framework, allows users to input data independently and obtain personalized diabetes risk assessments. This innovative platform serves as a powerful tool for early diabetes prevention and education, holding potential value in enhancing global diabetes management.

This paper describes an unmanned vehicle designed based on RTK precise positioning as well as visual image processing techniques. the whole system consists of a high-performance microcontroller (STM32H7), NVIDIA artificial intelligence board (Jetson nano), Gyroscope, GPS/RTK module, camera module, video camera and laptop. And based on the Qt Creator platform using C++ language to develop a set of data reception (navigation data, visual data), navigation path planning solution, human-computer interaction of the host computer software, and in the cloud platform to develop a set of user management system. The car through the fixed-speed cruise, according to the designated route to travel along the road fruit sample matching, judging fruit ripeness, fruit quality and GPS information, fruit information, etc. reported to the host computer for the agricultural intelligence unmanned detection program.

In response to the application requirements of low-power and high-performance MCU, this paper proposes a low-power microcontroller based on RISC-V instruction set and Dual stage pipeline architecture using a software hardware collaborative design method. It adopts a sequential Dual stage pipeline structure and supports configurable RV32I instruction set. The processor is designed using the comprehensible Verilog language and uses the ICB bus protocol as the interconnect bus to construct a System on Chip (SoC). In a simulation environment, the logic function of the microcontroller is verified by writing and inputting RV32I assembly instructions to test the program. After adding constraints under the Vivado synthesis tool, the processor RTL code is logically synthesized and the utilization of processor hardware resources is analyzed. Finally, the synthesized stream file is downloaded to the Artix7-100T FPGA chip development board. The verification results show that the designed microcontroller has correct logic function, uses low hardware overhead, and achieves relatively high performance indicators, which is suitable for high-performance with cost constraints Low power application areas.

With the advent of the Internet of Things era, the demand for low-power, low-cost, scalable, open, and flexible chips is increasing day by day. This paper aims to design a RISC-V processor with a three-level pipeline structure, and verify its performance and functionality through verification on FPGA. The design process includes steps such as RTL design and simulation testing, RTL synthesis, constraints, Place and Route, burning, running program development and downloading, and running the built program on the RISC-V processor burned on FPGA. The kernel and peripheral design adopts a three-level pipeline design, and the functions and tasks of each stage are clearly described. The correctness and compliance of the design have been verified through waveform simulation and official testing. The design was ultimately validated on FPGA and performance analysis was conducted, indicating that the design can achieve ideal results in practical applications.

Event stream has been used in various vision tasks due to the low latency and high dynamic range of event camera. However, because of the temporal dynamic of event stream, convolution neural networks (CNNs) are difficult to effectively extract features from event streams to achieve object tracking tasks. Besides, SNNs is suitable for processing data with temporal information because of its spiking delivery mechanism and membrane potential accumulation over time. In this work, we propose a Hybrid Neural Network (HNNet) to achieve effective event-based single object tracking tasks by combining the advantages of SNNs and Swin-Transformer. For higher feature expression ability of SNNs, we adopt the Swin-Transformer to extract features from sparse event stream. Then we use these features to modulate the threshold of SNNs neurons. What’s more, for improving tracking performance for both special and temporal features, a cross-modality fusion module is designed to fuse the two features extracted by the Swin-Transformer and SNNs. We conduct extensive experiments on three public event-based datasets (FE240, FE108, and VisEvent) and our tracker outperforms other trackers maximum at 1.1% and 6.8% in terms of area under curve (AUC) scores and precision rate respectively.

At present, ultrasonic cleaning technology has been widely used in various fields of industrial production. In order to gain a deeper understanding of the cleaning effect of ultrasonic cleaning technology on underwater accessories, this paper uses MATLAB to carry out ultrasonic field distribution and cavitation bubble dynamics simulation, and combines with the ultrasonic descaling mechanism, to design a set of ultrasonic cleaning device, and to assess the descaling effect of ultrasonic cleaning on Q345 alloy steel samples with different frequencies and different irradiation times. The experimental results show that when the input current is 6A and the irradiation time is 15 min, the descaling effect is best when the descaling rate reaches 0.58%.

In this paper, we present a fault classification approach that utilizes aeronautical text data by using improved Word2Vec and Convolutional Neural Network (CNN) models for fault diagnosis. The Word2Vec model is used for vectorization, and the fusion of word vector, character vector and N-gram method is used for text representation. In addition, the vector constructed based on the method are used as inputs to the CNN model, which is capable of extracting fault feature from the text data. The experimental results demonstrate that in comparison with other approaches, the method presented in this article has a better classification effect, the model has a certain degree of improvement in various indexes. In practice, it effectively decreases the manual reliance on fault text classification, and achieves the optimal performance indexes in accuracy, recall and f1-score, indicating the effectiveness of the method proposed in this paper.

Submersibles play a vital role in exploring shipwreck sites but face challenges like losing contact with the mother ship and mechanical failures. To address these, a submersible location prediction model was developed based on the turbulence collapse model, utilizing data on the Ionian Sea's currents, density, and terrain. The model incorporates an SST turbulence model, collapse processing, and a dynamic analysis of submersible velocity and displacement, with a state function describing the probability distribution of the submersible's location and an influence function for the next moment. This approach allows for defining uncertainty and suggests submersibles should periodically send their position, environmental parameters, and motion status to the mother ship, equipped with precision sonar, inertial navigation systems, environmental monitoring sensors, and communication devices.

Reproduced circuit boards (PCBs) are a crucial element in contemporary electronic goods, and their manufacturing process presents challenges for environmental control and has stringent criteria. Defects such as holes, mouse bites, open circuits, short circuits, stabs, and miscellaneous copper resulting from this can lead to product quality issues and subsequently affect the performance of electronic devices using the product. Currently mainstream target detection frameworks still face problems with small object detection when it comes to PCB surface defects. The study introduces an enhanced YOLOv8 algorithm designed for the real-time identification of surface irregularities on PCBs. Specifically, this includes introducing attention mechanism in the neck structure of YOLOv8 to enhance focus on key features, and replacing PANet feature pyramid structure with BiFPN structure to better capture boundary information. According to the experimental results, the improved model has achieved significant effects in detecting small target defects. Compared to the YOLOv8 algorithm, the accuracy of the improved algorithm has been increased by 1.5% and reached 95.4%. it can accurately perform defect detection tasks for printed circuit boards in industrial production and greatly enhance the accuracy of detecting various types of defects, demonstrating the effectiveness of the algorithm.

Lumbar intervertebral disc herniation is a common degenerative disease of the spine, and the application of deep learning technology to lumbar intervertebral disc herniation typing helps to improve the diagnostic efficiency of doctors and reduce the burden of doctors on tedious work. In this paper, a lumbar disc herniation typing method based on the ResNet50 network model is proposed. Firstly, to address the problem of a large-scale gap between different herniation types of lumbar disc foci, a multi-scale attention mechanism is introduced into the backbone network to improve the model's ability to capturing feature information of different scales; based on this, a double-branch structural model is constructed, with sagittal and axial positions as inputs to improve the robustness of the model by integrating information from different perspectives. information from different perspectives to improve the robustness and generalization ability of the model. The experimental results show that the method has a high accuracy of 88.1% in the test set.

In order to improve the dynamic obstacle avoidance ability and operational efficiency in unmanned vehicle path planning, a new algorithm incorporating the improved A* algorithm and DWA (dynamic window approach) is proposed and applied to indoor handling work. First, the valuation function is improved using the Euclidean distance and the heuristic function is dynamically weighted to reduce the expansion points, increase the real-time performance of the planning path, greatly reduce the search time, and improve the flexibility of the algorithm. Then, the path is optimized using Bessel curves to improve the smoothness. Finally, the dynamic window algorithm is integrated, and local path planning is carried out with the global optimal path guaranteed by the A* algorithm, so that it can dynamically avoid obstacles and improve path planning. Simulation experimental results show that the improved A* algorithm reduces the search nodes by 75%, the search time is reduced by 91.9%, the path length is still optimal, and the number of turns is reduced by 25%, while the fusion algorithm does not increase the path length and the search time by more than 3% after adding unknown obstacles. The final experiment verifies that the algorithm has good feasibility in real scenarios.

2D human pose estimation (HPE) has been a research focus of computer vision, 2D HPE baseline is one of the main studies. As the field of HPE continues to evolve, Vision Transformer Baselines have emerged as a significant area of interest, showing considerable potential in visual applications. However, accurate estimation remains a challenge in 2D HPE. This study introduces a novel approach named FANpose for 2D HPE in images. Building upon the top-tier VITpose Baselines, we innovate in two main aspects. Firstly, we employ fully attentional net- works to replace the vision transformer baseline model, thereby enhancing the model’s robustness. Secondly, we improve keypoint localization accuracy by replacing traditional Gaussian kernels with Laplacian kernels, thereby enhancing the model’s recognition precision. On the MS COCO dataset, our model achieves AP and AR scores that are respectively 0.4 and 0.6 higher than VITpose-B, and our model is 32M smaller in terms of parameters than VITpose-B.FANpose achieves satisfactory results in human pose estimation tasks, showcasing its immense potential for practical applications.

With the development of marine engineering and remote operated vehicles, underwater image enhancement is receiving more and more attention. In recent years, there have been many methods for underwater image enhancement. In particular, various convolutional neural networks (CNN) have been applied in the field of underwater image enhancement. However, the underwater imaging environment is complex and changeable. In the underwater environment, red light attenuates faster than blue-green light, and suspended impurities and various algae in the water will also affect the propagation of light. Therefore, the collected underwater light images often suffer from varying degrees of degradation. This makes it more challenging for CNN models to enhance images in different environments. This paper modifies the architecture of Water-Net, and improves the output of the confidence map. Experimental results on the Underwater Image Enhancement Benchmark (UIEB) show that our network achieves better results than other algorithms.