DUNet: Dense U-blocks network for fine-grained crack detection

Shibo Sheng, Hui Yin, Ying Yang, Ai-Xin Chong, Hua Huang

Abstract: At present, many progress has been made in crack detection methods based on deep neural networks. Compared to general cracks, fine-grained cracks are more difficult to detect not only due to their small and narrow shape, but also the various types of cluttered scenes. It is time-consuming to collect and label the samples of fine-grained cracks that makes most current data-driven models fail for poor generalization. To tackle these problems, a novel dense U-blocks network (DUNet) is proposed for fine-grained crack detection in this paper. Specifically, in order to preserve the integrity of accurate position information of the fine-grained cracks, a constant resolution U-block group (CRUG) is designed. Further, a dense feature connection strategy (DFCS) is proposed to enhance the information flow for better reuse of multi-scale features. DUNet achieves state-of-the-art performance on four fine-grained crack datasets, together on three general crack datasets with different environments. Moreover, we propose a lightweight version (DUNet-L) of DUNet for real-time practical applications with good accuracy and less parameters and computation.