AWC-ECRA-YOLO: A lightweight sonar defect detection network with adaptive wavelet convolution and enhanced cluster routing attention

This study presents AWC-ECRA-YOLO, a real-time sonar defect detection model designed to enhance the automation of sewer inspection in urban infrastructure systems. The proposed model addresses three major challenges in sonar imagery: low contrast, strong noise interference, and high computational cost. To this end, we introduce Adaptive Wavelet Convolution (AWC) for multi-frequency feature enhancement and background suppression, and Enhanced Cluster Routing Attention (ECRA) to improve global context modeling through prototype clustering and dynamic temperature scaling. Additionally, a novel Boundary Focusing Loss (BFLoss) increases sensitivity to small or ambiguous targets near decision boundaries.Extensive experiments on the Urban Sewer Sonar Defect Dataset (USSD) and the Underwater Acoustic Target Detection Dataset (UATD) demonstrate that AWC-ECRA-YOLO achieves superior performance, with mAP@50 gains of 0.82\% and 0.73\% over YOLOv11, respectively. By enabling fast, accurate, and automated defect identification in complex underwater environments, our method contributes directly to intelligent infrastructure inspection and supports automation in construction maintenance workflows.