An improved U-Net model for concrete crack detection
Chenglong Yu, Jianchao Du, Meng Li, Yunsong Li, Weibin Li
Abstract
Crack detection plays an important role in disease assessment of concrete buildings. However, factors such as complex background, irregular edge, and the real-time and accuracy requirement also make crack detection a challenging task. Aiming at the above challenges, an improved U-Net model for concrete crack detection is proposed, which has strong capability to extract the linear object, improving the performance in crack detection. The model is named Residual Linear Attention U-Net (RLAU-Net). There are three key measures in this paper. First, mirror padding the source image before convolution. Second, the multi-level features are obtained by aggregating the multi-scale features level by level. Third, strip pooling kernels are used to extract global contextual information, reducing information interference from the background. We tested the performance of RLAU-Net on our crack dataset, and the experimental results exhibited that it can improve the quantitative results of mean Intersection Over Union to 81.69%. In addition, F1 score has increased to, 78.21%, the Intersection Over Union of crack increased to 64.47%. We also compared the detect time-consuming of RLAU-Net and that of the original U-Net. Results demonstrate that the proposed model has a short processing time while maintaining a high detection accuracy for crack detection.