Crack Detecting by Recursive Attention U-Net

Crack detecting is a specific domain of semantic segmentation task for solving real-world applications such as pavement crack detection, bridge bottom crack inspection, solar cells or battery components defect detection. Different with general image with rich texture sematic information, road/bridge crack images are often in lack of semantic information which is vital for guiding segmentation. Thus, crack detecting task remains challenge in spite of great success of semantic segmentation neural networks. In this paper, we proposed a simple but powerful U-Net, named as "RAU-Net", to boost the crack detecting performance by recursive attention mechanism. First, the recursion residual block is used to extract low-level edge features. Then, attention mechanism is introduced for separating the saliency area and the irrelevant background area for accurate locating the cracks. Finally, recursive attention is equipped into U-Net architecture for efficient segmentation. Experimental results demonstrate that the proposed "RAU-Net" outperforms other state-of-the-art neural networks based crack detecting methods.