Abstract: The safe operation and maintenance of urban underground water supply and drainage systems are crucial for the sustainable development of society and the economy. The drifting capsule robot, as a new type of automated inspection tool, can effectively address the high costs and low efficiency of traditional methods. However, factors such as water flow disturbances inside the pipeline and the self-heating of the equipment can cause water mist interference in the images collected by the capsule robot, severely affecting the accuracy of defect identification. Therefore, this paper designs a lightweight defect detection network based on dual-branch feature fusion dehazing to improve the image quality and defect identification accuracy of low-quality capsule robot images. The dual-branch dehazing module adaptively extracts and fuses the spatial structure and spectral features of the image, enhancing the image's dehazing performance. The lightweight detection module, with YOLOv5 as the backbone, identifies and locates pipeline defect types and damaged areas in the dehazed images. Experimental results on the SFCRI dataset show that the dehazing module improves the SSIM and PSNR of the images by 0.271 and 24.04, respectively, and increases the average precision of defect identification by nearly 12%, achieving a recognition speed of 120.3 frame/s. The proposed lightweight defect detection network based on the dual-branch dehazing module can effectively achieve high-efficiency and low-cost defect detection for urban underground water supply and drainage pipelines.

Key words: pipeline defect detection, floating capsule robot, dual-branch defogging, deep learning