MRA-UNet: balancing speed and accuracy in road crack segmentation network

Road crack detection is an important step in urban road maintenance. Manual detection is both time-consuming and costly. Machine detection has gradually become the primary method in road maintenance due to the advantages of automation, stability and reliability. Therefore, this paper proposes a new UNet based on multi-residual attention (MRA-UNet), which is used in road crack segmentation. This paper proposes the multi-scale residual module which is used to capture crack information of different scales in the down-sampling path, and the dual-attention module is used to recover semantic information in the up-sampling path. The dual attention module in this paper is plug-and-play. At the same time, the MRA-UNet is evaluated on two public datasets, namely the CFD and EdmCrack600. Compared with some previous semantic segmentation networks, the proposed MRA-UNet achieved the good results of F 1 score, IoU, computational volume and inference speed, balancing the accuracy and speed of segmentation. Especially in terms of speed, this is more conducive to the practical application and ground deployment of the backorder, which can be applied to the road maintenance projects.