SP-CrackNet: serial–parallel network with boundary contrastive learning for real-time crack detection

Crack detectors based on deep learning have made tremendous progress compared to inefficient traditional inspections. However, missing thin local structures, blurred boundary segmentation and slow inference speed restrict the performance improvement of existing crack detectors. To this end, we propose a serial–parallel network (SP-CrackNet) for pixel-level real-time crack detection. Specifically, a serial–parallel feature extractor with global bottleneck blocks (GB Block) is proposed. Based on the serial–parallel structure, SPFE can enlarge the receiving field and capture the local information of thin cracks effectively, and the GB Block designed by us can ensure the continuity of the slender cracks is sensed with a lower computational cost. Moreover, a boundary contrastive learning scheme (BCL scheme) is designed to enhance the learning ability of SP-CrackNet for crack boundary features, thereby improving the segmentation accuracy of crack boundary regions. Extensive experiments on CFD and DeepCrack datasets show that SP-CrackNet outperforms the comparative methods while achieving real-time inference speed.