LSD-DETR: A Lightweight Transformer-Based Detector for Sewer-Pipeline Defects in Forward-Looking Sonar Images

Abstract

The regular inspection of underground sewer pipelines is essential for maintaining their functionality and supporting sustainable urban development. While optical sensors lack reliability, sonar imagery typically suffers from low contrast and severe noise, which is compounded by the constrained computational resources of remotely operated vehicles (ROVs). To address these limitations, this paper develops LSD-DETR, which is a lightweight Transformer-based detector for sewer pipe defects that balances accuracy and computational efficiency. First, Sonar-DualNet organizes CSP-based and HGNetV2-based blocks into a shared-stem heterogeneous backbone to reduce duplicated shallow computation while preserving complementary representations. Second, DR-Fusion performs content-adaptive modulation and aggregation of the two heterogeneous branches. In addition, the original CCFM and AIFI modules are replaced by a GLSA-enhanced BiFPN and the HiLo attention mechanism, respectively, as task-oriented adaptations for efficient cross-scale and local–global feature modeling. Experiments on the UPSD dataset show that LSD-DETR achieves an mAP50 of 89.6% and an mAP50:95 of 54.3%. Relative to RT-DETR-R18, these results represent improvements of 2.0 and 1.6 percentage points, respectively, while reducing the parameter count by 42.7% and FLOPs by 58.4%. The results demonstrate a favorable accuracy–complexity trade-off for forward-looking sonar defect detection in underground sewer pipelines.