Research on lightweight detection model of tunnel water leakage based on semantic segmentation

To solve the problems of large model parameters, slow detection speed, and poor anti-interference ability in complex backgrounds of traditional computer vision water leakage detection algorithms, a lightweight segmentation model SC-DeepLabV3+ based on DeepLabV3+ was proposed to realize the efficient detection of tunnel water leakage. First, the lightweight network S-Efficientnet, which integrates spatial and channel squeeze and excitation, was used as the backbone feature network. It can reduce model parameters and improve the speed of water leakage detection. The scSE attention mechanism compresses and enhances information in both spatial and channel dimensions to highlight the useful channels and suppress the useless channels, which in turn improves the detection accuracy of the model. Second, a high-level semantic feature refinement module Contact-Atrous Spatial Pyramid Pooling was designed. This module enhances the attention to the details of water leakage edges while considering the global contextual information by narrowing the expansion factor, adjusting the number of feature extraction branches, and fusing the depth separable convolution. The C-ASPP module processed and integrated semantic features from different scales to realize the mutual intermixing of the multi-scale semantic information and enhanced the anti-interference ability in complex environment. Finally, experiments were conducted on the constructed water leakage hybrid dataset. The results show that the S-Efficientnet backbone network greatly reduces the complexity of the model and enhances the focus on the water seepage area; the C-ASPP module improves the comprehensive utilization of multi-scale information; the intersection and merger ratio of the SC-DeepLabV3+ model reaches 90.17%, the size of the model is only 5.457 M, and the image processing speed is up to 89.525 f/s. Compared with existing mainstream semantic segmentation models, the SC-DeepLabV3+ model demonstrates significant superiority in terms of detection accuracy and segmentation speed. © 2024, Central South University Press. All rights reserved.