A physics-prior-guided algorithm for detecting surface cracks in underwater bridge structures

Underwater bridge crack detection remains challenging because underwater optical images are often degraded by light absorption and scattering, leading to color distortion, uneven illumination, low contrast, and complex background interference. In addition, annotated underwater crack samples are scarce. To address these issues, this paper proposes an underwater bridge crack detection method that integrates crack-related physical-prior extraction with the SegFormer framework. Five crack-related physical priors, including morphological continuity, texture abruptness, directional consistency, scale‑related response, and frequency-domain regularity, are quantified as interpretable physical features and fused to construct a multi-dimensional physical-prior representation, thereby improving crack segmentation accuracy. To alleviate sample scarcity and facilitate adaptation from terrestrial to underwater crack images, a two-stage transfer learning strategy is adopted. The model is first pre-trained on terrestrial crack images and then fine-tuned using limited underwater crack data. Experiments on the underwater crack dataset show that the proposed method achieves an mIoU of 90.54% and an mPA of 94.93%. Compared with the original SegFormer and DeepLabv3+, it improves mIoU by at least 5.35% and mPA by at least 3.52%. With a model size of 29.2 MB, it remains much smaller than DeepLabv3+ while achieving higher accuracy and better practical applicability.