Abstract This paper proposes the Bridge Crack Segmentation Network (BCS-Net), an end-to-end, lightweight segmentation framework designed for accurate and robust detection of complex cracks on concrete bridges. The network integrates an Adaptive Feature Selection Module (AFSM), which applies global pooling to enhance detail retention on small-scale feature maps and introduces Expectation-Maximization Attention (EMA) for efficient downsampling on large-scale maps. A Dynamic Snake Convolution (DSConv) is employed to capture elongated and curved crack patterns. Additionally, a combination of PixelShuffle upsampling and PixelUnshuffle downsampling ensures structural symmetry and effective feature transmission. Experimental results on both self-built (1,100 images) and public (348 images) datasets demonstrate that BCS-Net achieves the best balance between accuracy and efficiency, outperforming U-Net, DeepLabv3+, SegFormer, and PSPNet. It achieves F-measure/IoU scores of 89.44%/81.53% on the self-built dataset and 80.90%/68.83% on the public dataset, respectively, with an inference speed of 165.81 FPS (≈6 ms per image at 512×256 resolution).
Peng et al. (Sun,) studied this question.