An Efficient and Robust Model for Precise Surface Crack Segmentation on Concrete Bridge Structures

This paper proposes an efficient and robust model named bridge crack segmentation network (BCS-Net), tailored for high-precision segmentation of surface cracks on concrete bridges. BCS-Net adopts an encoder-decoder framework augmented by a multiscale feature extractor. A novel learnable hybrid sampler (LHS) is introduced to adaptively coordinate two complementary sampling strategies, reducing information loss from spatial compression and expansion while enhancing the preservation and reconstruction of crack details. Additionally, a composite enhancement module (CEM) is designed to strengthen the model’s focus on crack-relevant regions by synergistically integrating channel-wise and spatial attention mechanisms, effectively retaining critical structural cues while suppressing irrelevant features. Extensive experiments on both self-built and public data sets demonstrate that the proposed BCS-Net achieves an optimal trade-off between segmentation accuracy and computational efficiency, surpassing several mainstream segmentation models (U-Net, DeepLabv3+, SegFormer, HRNet-OCR, and Swin-Unet) and crack-specific models (CDU-Net and DTrC-Net). Specifically, BCS-Net attains F-measure scores of 89.53%, 86.31%, and 73.55%, and intersection-over-union (IOU) scores of 81.04%, 75.92%, and 58.16% on data sets comprising 1,000 self-built samples, 700 public samples, and 1124 CRACK500 samples, respectively. Moreover, BCS-Net delivers an inference speed of 13.84 frames per second (FPS), corresponding to a processing time of approximately 72 milliseconds per 1,024×512×1 resolution image, highlighting its strong potential for real-time deployment in automated bridge inspection scenarios.