Concrete Micro Crack Detection and Segmentation Using Adaptive U-Net

Abstract

The identification of concrete micro cracks has always been a difficult problem in crack detection, because the small cracks provide less effective detail information in the image, and the distribution is in disordered condition and irregular, complex environmental background. The existing semantic segmentation network based on deep learning can extract the crack location information to the pixel level, but micro cracks are too small that the proportion of the target pixel is very low, resulting in data imbalance which is the mainly reason that the small crack target is often ignored as the background target and can’t get the fine details. The accuracy of fracture identification is seriously affected. Pointing to this problem, this paper presents a small crack detection and segmentation method based on adaptive U-Net network. Aiming at the problem of uneven distribution of small crack sample data, firstly this method uses the oversampling strategy which expands the number of tiny cracks in the sample; at last due to the unbalanced characteristics of tiny cracks, the appropriate loss function is selected to adapt the U-Net which makes the model pay more attention to the tiny target. Experimental results show that the proposed method can effectively extract the features of small cracks, improve the similarity of clustering, enhance the semantic segmentation with powerful optimization performance and solve the contradiction between accuracy and speed. This method improves the robustness of small crack segmentation and reduces the risk of matching unsuitably, and can be more accurately detected and segmented than other semantic segmentation models.