Automated Pixel-Wise Pavement Crack Detection by Classification-Segmentation Networks

Pavement crack detection on pixel-levels is a high-profile application of computer vision and semantic segmentation. In this paper, a two-step convolutional neural network (CNN) method is proposed to detect crack-pixels from pavement pictures and to reduce time consumption. The method contains two main parts: CNN-1 for patch classification and CNN-2 for semantic segmentation. The first part chooses regions with a high probability to contain cracks and sends them to CNN-2 to get pixel-wise detection results. The CNN-2 cancels down-sampling to ensure the size of a feature map is fixed, so it is an end-to-end network. The proposed method and CrackNet-II are trained and tested on the same datasets, and the results show that compared with the pure-segmentation network, the two-step CNN method reduces the processing-time dramatically while the loss of accuracy is small.