A TransUNet Network With Dice Loss for Irregular Road Crack Segmentation

Road crack segmentation is of great significance for ensuring road safety, extending pavement lifespan, and reducing maintenance costs. However, current crack segmentation faces challenges such as difficulty in identifying slender cracks, strong interference from complex backgrounds, and diverse crack shapes, making it difficult to accurately segment slender cracks. To address these issues, this study proposes a series of improved U-Net segmentation schemes, including TransUNet, U-NetSA, and U-NetCA. Specifically, TransUNet demonstrates superior performance in enhancing global feature fusion and improving robustness. U-NetSA excels at handling the edges of irregular cracks; and U-NetCA significantly improves segmentation accuracy for tiny cracks and under complex backgrounds. Additionally, Dice Loss is introduced to address class imbalance, enhancing model convergence and accuracy. Experimental results on the Crack500 dataset show that TransUNet achieves a Dice coefficient of 81.63% and an accuracy of 98.62%, demonstrating the effectiveness and potential of the proposed methods for road crack segmentation.