Enhanced Accurate Segmentation of Bridge Surface Image Defects Based on Multi-scale Feature Fusion

Bridge defects will affect the normal operation of the bridge, and daily inspection and maintenance are critical to the service life of the bridge. At present, bridge defect detection has problems of intra-category inconsistency and inter-category indistinguishability. Therefore, we propose a bridge defect detection model MP-Unet, which solves the problem of low-level features with rich spatial resolution information but weak semantic information and object edge details. Fracture problem to improve the accuracy of bridge defect detection. The experimental results show that MP-Unet improves mIoU by about 3 percentage points to 78.56% compared with the Unet model, which is better than the commonly used segmentation models and can more accurately segment the bridge deck defect area.