Pixel-level Crack Detection using U-Net

In this paper, we proposed an automatic crack detection method based on deep convolutional neural network-U-Net [4]. Unlike existing machine learning based crack detection methods, we can process an image as a whole without patchifying, thanks to the encoder-decoder structure of U-Net. The segmentation result is output from the network as a whole, instead of aggregation from neighborhood patches. In addition, a new cost function based on distance transform is introduced to assign pixel-level weight according to the minimal distance to the ground truth segmentation. In experiments, we test the proposed method on two datasets of road crack images. The pixel-level segmentation accuracy is above 92% which outperforms other state-of-the-art methods significantly.