Automatic pavement crack segmentation using a generative adversarial network (GAN)-based convolutional neural network

Due to the increasing demand on road maintenance around the whole world, advanced techniques have been developed to automatically detect and segment pavement cracks. However, most of methods suffer from background noise or fail in fine crack segmentation. This paper proposes a generative adversarial network (GAN)-based neural network named CrackSegAN to segment pavement cracks automatically. The generator of CrackSegAN generates segmentation results, while the discriminator trains the generator adversarially. A joint loss function is proposed to optimize the generator with sufficient gradients and mitigate the high class imbalance in pavement crack images. Elastic deformation data augmentation method is applied to force CrackSegAN to learn the transformation invariance. The proposed CrackSegAN reaches an average F1 score of 0.9780 on CrackForest dataset and 0.8412 on Crack500 dataset. Ablation study shows that the most prominent difference is made by the proposed joint loss function which increases the average F1 score by 8.98% on CrackForest dataset. Besides, the comparison between using different data augmentation strategies validates the effectiveness of elastic deformation. Overall, the proposed CrackSegAN increases the F1 score by 1.91% on CrackForest dataset and 1.01% on Crack500 compared with state-of-the-art methods. Qualitatively, CrackSegAN is more robust to background noises and segments cracks with more details. Moreover, the test on field data proves a better generalizability of CrackSegAN on unseen background noises. •A generative adversarial network (GAN)-based convolutional neural network named CrackSegAN for pixel-level pavement crack segmentation.•Elastic deformation is applied to train CrackSegAN to learn the transform invariance.•A joint loss function is proposed to mitigate the high class imbalance.•CrackSegAN is more robust to background noise and segments more accurate crack boundaries compared with state-of-the-art methods.