Intelligent Defect Detection Method for Buried Sewage Pipelines Based on Wavelet Transform and Sonar Imaging

Abstract The primary objectives of buried sewer pipeline inspection are to ensure structural integrity, protect the ecological environment, and minimize economic losses. However, traditional optical imaging methods suffer from severe degradation in turbid underwater environments due to light attenuation, scattering, and absorption, making them inadequate for practical structural health monitoring. To address these challenges, this paper proposes an intelligent defect detection approach based on sonar imaging, termed discrete wavelet transform–enhanced You Only Look Once (DWT-YOLO). Unlike existing YOLO-based methods that apply frequency-domain processing or spatial attention mechanisms independently, the proposed approach introduces a tightly coupled frequency–spatial feature enhancement strategy. By deeply embedding discrete wavelet transform into the feature aggregation process, frequency-domain information is enabled to interact with spatial features through collaborative modeling and end-to-end optimization, thereby overcoming the limitations of conventional spatial-domain methods under speckle noise, low contrast, and texture degradation conditions. Experimental results showed that DWT-YOLO achieves a segmentation mean average precision (mAP) at an intersection over union (IoU) threshold of 0.5 (mAP 0.5 ) of 92.5% and mAP averaged over intersection over union (IoU) thresholds from 0.5 to 0.95 (mAP 0.5-0.95 ) of 46.9% on the buried sewer pipeline sonar image data set, improving by approximately 4% and 1.5%, respectively, compared with the baseline YOLOv11, while slightly reducing the number of parameters by 0.05 million parameters and requiring 9.5 Giga floating-point operations (GFLOPs) of computation. Furthermore, the proposed method can significantly enhance the efficiency with which engineers identify target defects from sonar images, thereby providing strong support for downstream engineering tasks.