Pixel-level pavement crack segmentation with encoder-decoder network

Crack detection is important to pavement condition surveys. The convolutional neural network (CNN) is one of the most powerful tools in computer vision. However, pixel-perfect crack segmentation based on CNNs is still challenging. This paper proposes an encoder-decoder network (EDNet) for crack segmentation to overcome the quantity imbalance between crack and non-crack pixels, which causes many false-negative errors. The decoder of the proposed EDNet is an autoencoder and self-encodes the ground-truth image to corresponding feature maps that are completely abstract, resulting in significantly reduced quantity imbalance between crack and non-crack pixels. Therefore, instead of fitting crack images directly with ground-truth images, EDNet's encoder fits crack images with corresponding feature maps to overcome the quantity imbalance problem. EDNet achieves overall F1-scores of 97.80% and 97.82% on 3D pavement images and the CrackForest dataset, respectively. Experimental results show that EDNet outperforms other state-of-the-art models.