Lightweight Semantic Segmentation of Steel Bridge Coating Corrosion States Using Scale-Matched Detail Injection and Efficient Channel Attention

Visible corrosion on coated steel bridge surfaces is manifested through irregular changes in color, texture, apparent roughness, and boundary morphology. RGB inspection images can record these surface appearance features, but they do not directly measure coating thickness, coating adhesion, corrosion depth, or residual structural capacity. This study therefore focuses on the pixel-level segmentation of four visible corrosion condition states—Good, Fair, Poor, and Severe—and proposes a lightweight semantic segmentation network, termed DE-YOLO26n-sem. Built on the YOLO26n semantic segmentation baseline, the network uses a stride-2 convolution to align shallow P2/4 features with the P3/8 feature scale before injecting them into the main segmentation branch. This design helps preserve corrosion boundaries and fine surface details weakened by downsampling. An efficient channel attention module then adaptively recalibrates color and texture responses without reducing the channel dimension. Training combines weighted cross-entropy, Dice loss, and auxiliary supervision. The method is evaluated on 440 pixel-annotated images from the Corrosion Condition State Semantic Segmentation Dataset v2 using a five-fold nested procedure. In each outer iteration, one-fold is reserved exclusively for testing, while architecture screening and checkpoint selection are conducted only within the remaining development data. A6, which combines scale-matched detail injection and efficient channel attention, is selected in four of the five inner screening procedures, while A4 is selected in the remaining iteration. Across the five independent outer test folds, the nested-selected configurations, comprising A6 in four folds and A4 in one fold, achieve an mIoU of 41.52% ± 0.66%, compared with 40.76% ± 0.85% for the baseline. Severe-class F1 increases from 14.10% ± 5.60% to 18.72% ± 6.20%, while pixel accuracy changes from 82.61% ± 0.71% to 82.34% ± 0.65%. In the single-seed efficiency evaluation, the model has 1.694 million parameters and requires 1.896 GFLOPs, with a forward-pass latency of 4.07 ms for a 384 × 384 image on an RTX 4070 SUPER GPU. Ablation experiments show that a direct P2 output branch and CARAFE do not provide consistent improvements; neither is included in the final network. The resulting pixel-level maps characterize the spatial distribution of visible corrosion appearances and can support digital inspection records and the prioritization of follow-up examinations. The resulting pixel-level condition maps can support digital inspection records and help identify image regions requiring follow-up examination. The predicted classes should not be interpreted as direct measurements of coating degradation, corrosion depth, section loss, residual load-carrying capacity, or remaining service life, all of which require additional material or structural measurements.