RobustFlow: An unsupervised paradigm toward real-world wear detection and segmentation with normalizing flow

Abstract

Three main challenges in industrial wear detection that limited data-availability and time-consuming annotations, small areas of initial wear, and sensitivity to variable light, have impeded the real-world applications of deep learning-based methods. To this end, we propose RobustFlow, an unsupervised method based on the normalizing flow and attention mechanism. In our work, only the wear-free images are required for training, and then the trained model can be employed to detect and segment wear. Extensive experiments have demonstrated that RobustFlow can achieve predominant robustness in real-world wear detection and segmentation, especially for wear with small regions and variable light. Overall, our work provides a promising paradigm for wear detection and segmentation in real-world industry.