Bridge Crack Detection Method Combining CNN and Trans-former

Regular inspection of bridge cracks is a crucial but labour-intensive task, which is aimed at the problems of complex background noise interference, fine crack features, and the bridge cracks are composed of multiple materials, with a large number of misdetections and omissions, which This paper proposes a novel crack segmentation model by integrating ResNet-50 and Swin Transformer V2 as the encoder. A feature fusion module is designed to efficiently detect cracks in the bridge. feature fusion module is designed to efficiently aggregate intermediate layer features from both branch encoders and connect them to the corresponding A feature fusion module is designed to efficiently aggregate intermediate layer features from both branch encoders and connect them to the corresponding layer decoder, enhancing the detailed feature information. Sub-pixel convolution is introduced in the decoder for up-sampling to better restore the details of the cracks. Sub-pixel convolution is introduced in the decoder for up-sampling to better restore the details of the cracks. Finally, the model is trained using a combined loss function to alleviate the crack sample imbalance problem. results demonstrate that on the public Bridge Crack Library (BCL) dataset, the proposed model achieves precision, F1-score, and mean Intersection over Union (mIoU) of the cracks. Union (mIoU) of 77.84%, 79.17%, and 82.07% respectively. Compared to ten other models, it achieves the best performance, effectively mitigating false detections and missed detections, and significantly improving the detection accuracy of fine cracks.