A multi-scale frequency attention fusion-based framework for structural defect segmentation in sonar imaging of buried sewage pipelines

In-service buried sewage pipelines may develop structural defects due to long-distance transport, changes in soil pressure, ground settlement, or material ageing. To reduce labour and resource consumption, pipeline maintenance personnel aim to detect defects without draining the pipeline, even under fully filled or high-water-level conditions. Sonar inspection, capable of penetrating water and sediment, has proven to be an effective approach. However, existing sonar-based defect detection methods are limited in automation and precise classification, often relying on manual analysis, which hinders their applicability in large-scale urban pipeline networks. To address this, this paper proposes a structural defect segmentation method for sonar imaging of buried sewage pipelines. The method proposes a multi-scale discrete cosine transform attention (MultiScaleDCTAttention) that fuses spatial and frequency-domain features to enhance multi-scale target and image detail perception. Additionally, a sonar feature fusion module (SFFM) is designed, integrating spatial attention and channel attention to strengthen feature representation from multiple dimensions. Experimental results on both buried sewage pipeline and public forward-looking sonar (FLS) datasets show that the proposed method achieves mAP0.5 of 0.916 and 0.825, and mAP0.5–0.95 of 0.472 and 0.502, respectively, significantly outperforming existing methods in segmentation accuracy and environmental adaptability.