Self-Supervised Iterative Refinement for Anomaly Detection in Industrial Quality Control

TL;DR

This work tackles the challenge of anomaly detection in industrial quality control under noisy and sparsely labeled data. It introduces the Self-Supervised Iterative Refinement Process (IRP), which combines self-supervised learning with a probabilistic scoring framework and a dynamic, median-based threshold to iteratively prune misleading samples and retrain models. Across KSDD2 and MVTec-AD benchmarks, IRP consistently outperforms traditional methods like DifferNet and One Shot Removal, particularly in high-noise scenarios, demonstrating improved AUROC and robustness. The approach also provides insight into adaptive data refinement and anomaly localization, offering a practical pathway to more reliable industrial defect detection and quality assurance.

Abstract

This study introduces the Iterative Refinement Process (IRP), a robust anomaly detection methodology designed for high-stakes industrial quality control. The IRP enhances defect detection accuracy through a cyclic data refinement strategy, iteratively removing misleading data points to improve model performance and robustness. We validate the IRP's effectiveness using two benchmark datasets, Kolektor SDD2 (KSDD2) and MVTec AD, covering a wide range of industrial products and defect types. Our experimental results demonstrate that the IRP consistently outperforms traditional anomaly detection models, particularly in environments with high noise levels. This study highlights the IRP's potential to significantly enhance anomaly detection processes in industrial settings, effectively managing the challenges of sparse and noisy data.

Figures (5)

• Figure 1: Typical impact of noisy data in training an anomaly detection model.
• Figure 2: Illustrative schematic of the Iterative Refinement Process, demonstrating the independent cycle of training, validation, retraining, and testing.
• Figure 3: Figure (a) displays the performance trends of various models on the KSDD2 dataset, while figure (b) shows results for the MVTec-AD dataset, across different noise levels. The graphs illustrate the models' performance before and after applying the IPR model: the blue line represents the IPR model, the green line denotes the OSR model, and the yellow line illustrates the performance of the vanilla model.
• Figure 4: AUROC performance of the IRP across various classes of the MVTec-AD dataset. Each sub-figure demonstrates the defect detection efficacy for a different class, illustrating the robustness and adaptability of the IRP in handling diverse defect characteristics.
• Figure 5: A visual representation of good (outlined in green) and defective (outlined in red) samples under different noise levels. Defects are marked with red circles, illustrating the distribution of flaws within the samples.