Multiscale feature fusion-based pavement crack detection using TransUNet

This study suggests an enhanced and optimized detection approach based on TransUNet to address the problems of poor crack recognition accuracy and imprecise edge information extraction in complicated backgrounds. The detection model is built upon TransUNet, with the inclusion of Coordinate Attention (CA) module in the convolutional neural network (CNN) section of the decoder. This modification enables the model to more accurately capture the spatial features and channel-level dependencies of the image. Additionally, Dense Atrous Spatial Pyramid Pooling (ASPP) module is incorporated at the three skip connection points, integrating and utilizing feature information at different scales to obtain feature maps that cover more semantic information. On the self-constructed dataset, the model shows improvements in average Intersection over Union (mIoU) and F1-score of 8.06% and 5.76%, respectively, as compared to the baseline network. Furthermore, when compared to other mainstream segmentation models on both the self-built dataset and public crack datasets CRACK500 and CFD, the proposed model outperforms six other models across all evaluation metrics. These results indicate that the proposed method effectively addresses the challenges of pavement crack detection in complex backgrounds.