An identification and classification framework for bridge cracks using transposed convolution-CBAM-Segformer with transfer learning

Bridge cracks seriously affect structural bearing capacity, durability and traffic safety, making accurate and efficient crack detection essential for bridge maintenance. To address the problems of feature degradation and upsampling information loss in existing segmentation models for bridge crack detection, this paper proposes an improved Segformer-based bridge crack identification and classification model. The model integrates the Convolutional Block Attention Module (CBAM) to enhance effective crack features and suppress background interference, and replaces bilinear interpolation with transposed convolution to reduce detail loss during upsampling. Meanwhile, a difficult-easy sample partitioning strategy based on crack pixel proportion is adopted to alleviate extreme pixel imbalance, and transfer learning is introduced to improve training convergence and small-sample generalization ability. Comparative experiments with mainstream segmentation models show that the proposed method achieves superior performance, with MIoU, Precision, and Recall reaching 92.14%, 95.75%, and 93.59%, respectively, which outperforms the MIoU, Precision, and Recall of U-Net, PSPNet and DeepLabV3 +, that is, 85.34%, 89.05%, 87.71%; 80.14%, 83.88%, 82.29%; 75.38%, 79.80%, 78.24%. The results demonstrate that the proposed model can effectively and stably implement accurate identification and material-based classification of bridge cracks under complex and small-sample scenarios.