End-to-end semi-supervised deep learning model for surface crack detection of infrastructures

Surface crack detection is essential for evaluating the safety and performance of civil infrastructures, automated inspections are beneficial in providing objective results. Deep neural network-based segmentation methods have demonstrated promising potential this purpose. However, majority these fully supervised, requiring extensive manual labeling at pixel level, which a vital but time-consuming expensive task. In paper, we propose novel semi-supervised learning model detection. The proposed employs modified U-Net, has half parameters original U-Net network to detect surface cracks. Comparison using 20 epochs shows that requires only 15% training time traditional U-net, improves accuracy by 20% upwards. On basis, (modified U-Net) trained based on an updated strategy. At each stage, predicts segments unlabeled data images. new strategy updating datasets allows be with limited labeled image data. To evaluate method, comprehensive consisting DeepCrack, Crack500 those open public, expanded dataset containing 2068 images concrete bridge our independent labels, used train test method. Results show method achieved quite approaching accuracies established supervised models multiple indexes, however, requirement reduces 40%.