Threshold Estimation-Assisted Unsupervised Patch-Wise Model for Industrial Inspection of Anomaly

The smart manufacturing revolution, driven by advanced technologies such as Artificial Intelligence and Internet of Things, is rapidly reshaping the world economy. One critical component of this revolution is industrial inspection and image analysis for anomalies, which are now being boosted by deep learning techniques capable of extracting latent features for high-quality defect detection. However, most existing unsupervised approaches combining deep models and outlier detection models pay little attention to the critical issue of threshold and relevant estimation, which is crucial for practical deployments. In this work, we propose a threshold estimation-assisted patch-wise approach that involves a core set selection of representative nominal patch features, anomaly scoring through comparing inputs and the core set, and a permutation-assisted threshold estimation. Our proposed model is technically practical for industrial deployment and addresses the cold-start problem with reduced labor for image annotation. To enhance the selection and scoring process, we also introduce novel position encoding and neighbor-assisted schemes. Our approach outperforms popular state-of-the-art baselines in extensive experiments involving images from different industry domains, demonstrating its practical effectiveness.