WT-DETR: A Lightweight Model for Defect Detection in Buried Sewer Pipelines Based on Frequency–Domain Feature Processing

Abstract

Advancements in artificial intelligence have provided promising solutions for the detection and classification of internal defects in buried sewer pipelines. However, deep learning methods in this domain still face challenges, particularly regarding limited detection accuracy for small defects and excessively large model sizes. This study proposes a methodology that utilizes sonar images and defect frequency-domain features for the accurate identification and classification of pipeline defects. First, raw sonar data collected and processed from field scans are analyzed and filtered. Subsequently, deep learning algorithms are trained and evaluated on the extracted features to recognize and classify pipeline defects. To capture low-frequency structural details, a feature extraction module termed re‑parameterized wavelet‑convolution (RepWT) is designed. Additionally, two distinct feature processing architectures are fused to reduce the computational overhead for deployment. Finally, by integrating the proposed inner-minimum‑point‑distance IoU (MPDIoU) loss function, a neural network model named wavelet convolution-detection transformer (WT-DETR) is developed for sonar-based pipeline defect recognition. The performance of WT-DETR is compared against several state-of-the-art models. Experimental results indicate that WT-DETR achieves highly competitive performance for buried sewer pipeline defect detection, reaching a precision of 88.5%. These results align with theoretical expectations, demonstrating the effectiveness of the WT-DETR model in buried sewer pipeline defect detection.