MGSCA: Multiscale Gradient Selection-Based Coupled-Hypersphere Adaptation for Unsupervised Pavement Defect Detection

The pavement defect detection is challenging due to the diverse defects and their unpredictable formations. Current methods often struggle to perform well in situations with complex pavement backgrounds and weak textures, a problem that arises from the interference of various pavement information. We proposed an unsupervised pavement defect detection method with a multiscale gradient selection-based coupled-hypersphere adaptation (MGSCA) to alleviate these challenges. Specifically, the proposed method first constructs a lightweight and representative feature-capturing memory bank to effectively capture features of complex pavement backgrounds and reduce the inference time. Utilizing these representative features as hypersphere centers, we propose a hypersphere-coupled adaptive module that trains an axis-based self-attention descriptor to reduce the data bias of the pretrained network and improve the defect detection from the nondefective pavement backgrounds. On the Crack500 dataset, the proposed method achieved 0.930 area under the ROC curve (AUROC) in defect detection. For defect localization, the proposed method achieved 0.925 AUROC, 0.791 per-region-overlap (PRO), and 0.744 DICE. Compared with several state-of-the-art methods, the results on three datasets demonstrate the effectiveness of our proposed method in different scenarios.