Study of Overlap-Preserving Multi-Label Semantic Segmentation for Visual Bridge Inspection

Reliable bridge defect segmentation is important for automated visual inspection of in-service bridges. However, segmentation remains challenging because of class imbalance, heterogeneous defect appearances, and overlapping annotations between damage classes and structural components. This study examines multi-label semantic segmentation on DACL10K v2 by representing its 19 classes as independent binary targets. An annotation audit shows that approximately 15.8% of annotated training pixels carry more than one label. The 6935 official-training images were divided into fixed internal-training and internal-validation subsets for model development, and the configurations selected on the internal split were subsequently assessed on the 975-image official validation set. DeepLabV3+, SegFormer, and Mask2Former are compared under a common protocol, and the strongest baseline is subsequently used to evaluate loss settings and feature-enhancement modules. The module comparisons and branch ablations lead to a two-branch Local–Global Context Refinement (LGCR) configuration. Across three random seeds, SegFormer and LGCR achieve internal-validation mIoU values of 38.93 0.48% and 39.50 0.15%, respectively. On the official validation set, their corresponding mIoU values are 37.38 0.17% and 37.44 0.88%. Across the three seeds, the mean LGCR–SegFormer difference on the official validation set was 0.06 percentage points, and the direction of the difference varied by seed. These results do not demonstrate a consistent or substantial advantage for LGCR. The study provides comparative evidence on architecture, loss rebalancing, and feature refinement choices under the multi-label annotation scheme of DACL10K v2.