Crack U-Net:Towards High Quality Pavement Crack Detection

• Published in: Ji suan ji ke xue Vol. 49; no. 1; pp. 204 - 211
• Main Authors: Zhu, Yi-fan, Wang, Hai-tao, Li, Ke, Wu, He-jun
• Format: Journal Article
• Language: Chinese
• Published: Chongqing Guojia Kexue Jishu Bu 01.01.2022; Editorial office of Computer Science
• Subjects: Accuracy; Coders; Convolution; Datasets; Electronic devices; Encoders-Decoders; Feature extraction; Flaw detection; Image segmentation; Pavements; road pavement|crack detection|deep learning|image segmentation; Scale models; Vehicle safety
• ISSN: 1002-137X
• DOI: 10.11896/jsjkx.210100128

Pavement cracks have a great potential threat to driving safety, and the previous manual detection methods are not efficient. The existing crack detection methods have low model generalization ability, poor crack segmentation ability and low efficiency in complex backgrounds. In order to solve these problems , this paper proposes a new improved network structure Crack U-Net based on encoder-decoder structure, the purpose is to improve the model generalization and detection accuracy of pavement crack detection. First, Crack U-Net is enhanced with dense connection structure The network U-Net model based on encoder-decoder is proposed, which improves the utilization of feature information at each layer of the network and enhances the robustness of the model on the basis of the previous structure. Secondly, Crack U-Net uses residual blocks and The Crack U-block composed of mini-U is used as the basic convolution module of the network. Compared with the traditional double-layer convolution layer, Crack U-block can