Reproduce “Robust Anomaly Map Assisted Multiple Defect Detection with Supervised Classification Techniques” on OpenResearch

About this paper

Jože M. Rožanec, Patrik Zajec, Spyros Theodoropoulos, Erik Koehorst, Blaž Fortuna, Dunja Mladenić

Researchers integrated anomaly maps, generated by an unsupervised defect detection model, as additional input features for supervised machine learning models to improve visual defect detection and classification. This enhanced approach demonstrates higher recall rates and increased robustness against class imbalance in real-world manufacturing datasets.

Key insights

• Anomaly maps provide complementary spatial information that, when combined with raw image features, consistently improves both binary and multiclass defect classification performance.
• Anomaly maps alone prove highly effective for binary defect detection, often surpassing image-only inputs in identifying the presence of a defect.
• Integrating anomaly maps significantly enhances the robustness of supervised models, allowing them to maintain higher performance even when trained with severely imbalanced datasets.

Summary by alphaXiv.