A two-stream crack detection model based on SegNet

Road crack detection methods based on deep learning need to extract features from images by stepwise down-sampling, while down-sampling will lose the information of edges. However, edge information is crucial for slender targets such as cracks. This paper proposes a multi-task deep learning model consisting of two branches for road image crack segmentation. The first branch is the semantic segmentation branch based on SegNet. The second branch is the edge detection branch which learns the edge of cracks to supplement the detailed information lost in feature extraction of the semantic segmentation branch to achieve better segmentation results. At the same time, the proposed model includes edge attention modules to guide the model to pay more attention to details such as edges and ignore irrelevant information in the training process. Through comparative experiments, the proposed model is superior to other mainstream semantic segmentation models in pixel accuracy, mean intersection of union, and frequency-weight intersection of union.