2.1

Model framework

This paper proposes two improvements to the YOLOv8 algorithm, 1) the introduction of the SCConv module, which is able to process both spatial and channel information of the image, and can better capture subtle defect features without bringing additional computational burden. 2) To improve the visual model’s feature extraction capabilities, the MSDA module integrates cavity convolution with a multi-scale self-attention mechanism. The improved YOLOv8 framework is illustrated in Figure 1 and is composed of five main parts: input preprocessing, trunk, neck, head, and output.

Figure 1.

Improvements to the YOLOv8 framework.

The algorithm uses a series of initial convolution and downsampling to extract the image base features, and SCConv module is applied to feature maps ranging from 160x160 to 80x80, to better capture the spatial structure and color depth information at different scales when dealing with defective images of underground pipelines. MSDA module generates query, key, and value features by linear projection and processes the features in the head using multi-scale null convolution, then fuses the feature maps at different scales.

2.2

Description of SCConv module

SCConv is a highly efficient convolutional module designed to address spatial and channel redundancies within CNN. By minimizing unnecessary information in both dimensions, SCConv not only reduces the computational burden but also enhances the overall performance of the network. This is achieved through an optimized feature extraction process that makes more effective use of computational resources. The module is composed of two core components: the Spatial Reconfiguration Unit (SRU), which focuses on reorganizing spatial information, and the Channel Reconfiguration Unit (CRU), which refines feature representation across channels. Together, these units enable more precise and efficient feature extraction, leading to improved network efficiency and accuracy.

SRU is part of the SCConv module and is responsible for reducing redundant features in the spatial dimension. This part receives the input features and processes them through the following steps: First, the input features are normalized to reduce the scale difference between different feature maps. Second, the features are weighted by a series of weights w₁,w₂,…,w_c, These weights are computed from the channels of the features r₁,r₂,…,r_c after normalization and nonlinear activation functions. Finally, the weighted features are partitioned into two parts and , These two parts are then each transformed and then reconstructed by addition to get the spatially refined feature X^w. SRU configuration is shown in Figure 2.

Figure 2.

Spatial Reconfiguration Unit (SRU) configuration.

The configuration of the channel reconstruction unit (CRU) is illustrated in Figure 3, which is designed to reduce the channel redundancy of the convolutional neural network features. The CRU further operates on the features that have been processed by the SRU, which starts with the spatially refined features X^w, splitting X^w into two parts. different 1x1 convolutions through different scaled α and (1 – α) paths, The two parts of the features are further transformed by Group wise convolution (GWC) and Point wise convolution (PWC), and two transformed convolved features Y₁ and Y₂ are pooled and SoftMax weighted fused to form the final channel refined feature Y.

Figure 3.

Channel Reconfiguration Unit (CRU) configuration.

2.3

Description of MSDA Module

The working principle of the multiscale dilated attention MSDA is shown in Figure 4. Within the MSDA module, the feature map channels are first segmented into several heads. Each of these heads utilizes a distinct dilation rate to apply self-attention, ensuring that various scales of information are effectively processed across the different channel segments. These operations are performed between colored blocks within a window surrounding a red query block, and the example in the figure shows three different dilation rates that correspond to different receptive field sizes (3x3, 5x5, 7x7). The self-attention manipulation of each head targets its corresponding dilation rate and receptive field. The model efficiently captures features across multiple scales, which are then combined and passed through a linear layer to aggregate these features. This multi-scale design enhances the model’s ability to interpret the image by considering various levels of detail, leading to a more comprehensive understanding of the image’s content. This design allows the model to understand the image at different scales, thus improving the overall understanding of the image content. With this approach, MSDA not only captures local details, but also senses the contextual information of a wider region, enhancing the expressive power of the model.

Figure 4.

Multi-scale dilated attention (MSDA) structure.

3.1

Dataset

The data used in the experiment were collected using CCTV equipment in a city in southern China, classified in accordance with the industry standard drainage pipe inspection and assessment technology (CJJ 181-2012), and a total of 19,039 images of underground drainage pipe defects using manual labeling of defect locations. In response to the presence of an imbalance in the number of categories in the dataset, data enhancement was performed on the categories with fewer categories in the dataset, a sum of 55,915 images were finally generated after adjustment. The dataset was divided into training, val and test sets in the ratio of 7:2:1, where 39,140 images were used as training set samples, 11,183 images were used as val set samples and 5,592 images were used as test samples. The dataset contains both structural and functional defects. Structural defects are crack, deformation, corrosion, misalignment, fluctuation, disjoint, leakage, penetratio, detachment, side-branch, functional defects are deposit, encrustation, scum, roots, obstacle, and broken-wall. A total of sixteen types cover the types of drain defects.

Figure 5.

Sample Drainage Pipeline Data.

3.2

Experimental platform

To guarantee the reliability and consistency of the experimental outcomes, all experiments in this study were carried out on a designated computing platform. The specific details of the experimental environment, including hardware and software configurations, are outlined in Table 1. This setup was carefully chosen to ensure that the results could be accurately replicated under consistent conditions.

Table 1.

Hardware and software configurations.

3.3

Metrics

In order to evaluate the improved YOLOv8 algorithm network model detection performance, the model results were evaluated using Precision (P), Recall (R), Average Precision (AP), Mean Average Precision (mAP) as evaluation metrics. The relevant formulas are shown below:

In the above equation, TP represents the number of true positive samples, which refers to instances where the model correctly identifies positive cases. FP stands for false positives, indicating the number of negative samples that the model mistakenly classifies as positive. FN refers to false negatives, which are negative samples that the model incorrectly classifies as positive. Lastly, C denotes the number of different defect types considered in the classification process.

3.4

Experimental results

All the input images in the experiment were uniformly resized to 640×640. the training process used a configuration with an initial learning rate of 0.01, and the optimizer chose the Stochastic Gradient Descent (SGD) algorithm. The batch size was set to 8 and the number of training epochs was set to 350. Figure 6 shows the comparison of the PR curves of the original model and the improved model. Obviously, the area under the PR curve of the improved model is larger, which indicates that its performance is significantly better than that of the original model.

Figure 6.

Comparison between before and after PR curves

3.5

Comparison experiment

To provide a more comprehensive evaluation of both the accuracy and efficiency of the improved model, we also conducted comparison experiments using several state-of-the-art object detection algorithms, including YOLOv8, YOLOv10, and RT-DETR. The results of these experiments, which are aimed at benchmarking the performance of the improved model against these advanced detection methods, are presented in Table 2 for detailed comparison and analysis. This allows for a clearer understanding of the model’s strengths in relation to other leading algorithms in the field.

Table 2.

Comparison results.

The experimental data indicate that the improved model achieves the optimal performance in all the metrics. Precision rate is 92.6%, recall rate is 89.9%, and it reaches 94.4% and 80.2% at mAP@0.5 and mAP@0.5:0.95 respectively, which are better than other compared models.