• AWC-ECRA-YOLO is proposed for real-time sewer sonar defect object detection • AWC extracts multi-frequency features to boost contrast and suppress noise • ECRA applies clustering and dynamic temperature to cut cost and enhance context • BFLoss improves detection of small and boundary-ambiguous sonar defects • USSD and UATD results surpass YOLOv11 in mAP with real-time speed Urban sewer inspection is essential for maintaining the safety and reliability of water and wastewater infrastructure, yet automated assessment remains challenging in water-filled and turbid pipelines where optical sensors are unreliable and sonar images are often low-contrast and noise-contaminated. This study presents AWC-ECRA-YOLO, a real-time sonar-based defect detection model designed to improve the automation of sewer inspection. The proposed model addresses three major challenges in sonar imagery: low contrast, severe noise interference, and high computational cost. Specifically, Adaptive Wavelet Convolution (AWC) is introduced for adaptive multi-frequency feature extraction and background suppression, while Enhanced Cluster Routing Attention (ECRA) improves global context modeling via prototype clustering and dynamic temperature control. In addition, a Boundary Focusing Loss (BFLoss) enhances sensitivity to small or ambiguous defects near decision boundaries. Experiments on the Urban Sewer Sonar Defect Dataset (USSD) and the Underwater Acoustic Object Detection Dataset (UATD) demonstrate that AWC-ECRA-YOLO achieves state-of-the-art performance, with mAP50 gains of 0.82% and 0.73% over YOLOv11, respectively. On USSD, the evaluated sewer defect categories include Disconnection, Misalignment, Corrosion, Obstruction, Penetration, Scaling, Sedimentation, Spalling, WellChamber, and Other. These results indicate that the proposed method enables fast and accurate defect detection in complex underwater environments, supporting efficient and scalable sewer infrastructure inspection.
Ge et al. (Sun,) studied this question.