Bridge Crack Segmentation Algorithm Based on Improved U-Net

This paper introduces an enhanced U-Net network designed to extract fine crack features against complex backgrounds and balance crack pixel categories for bridge crack segmentation. Initially, the model integrates an ECA (Efficient Channel Attention) module with the original U-Net to dynamically calibrate channel features. This enhances the detection of crack-related features and suppresses irrelevant textures. The architecture is further enhanced by a DSConv (Dynamic Snake Convolution) layer, which the employs learnable offset vectors to adaptively deform convolutional kernels. This allows for precise mapping along the contours of curvilinear cracks. Finally, the model utilizes a weighted combination loss function to achieve a balance between pixel-level classification and global shape consistency, effectively addressing gradient conflicts in scenarios with class imbalance. This study’s algorithm was validated for crack segmentation using a custom dataset. Experimental results show that it outperforms other algorithms in accuracy and improves upon the original U-Net across all metrics, with mIoU (Mean Intersection over Union) and F1-Score increasing by 1.23% and 1.89%, respectively.