Abstract
Pavement crack segmentation is a critical task for road safety but manual crack segmentation is extremely time consuming. In order to improve the segmentation effect of asphalt pavement cracks, we propose a pavement crack semantic segmentation network based on the improved U-Net model with the atrous convolution and attention mechanism. ResNet50 is selected as the backbone to extract feature of asphalt pavement cracks. The module of Dense Atrous Convolution (DAC) and Attention Mechanism for Crack(ACM) are integrated between the encoding part and the decoding part to obtain more crack details and global contexts, and make the crack location more accurate. The experimental results on the public dataset GAPs384 and the self-collected dataset shows that the mIOU, R and F1 score of the proposed ACAU-Net is better than other methods, which verifies the feasibility of the network.