Advancing Quality Control in Painted Metal Parts: A Novel Dataset and a Defect Detection Method Based on Synthetic Anomalies

Abstract

Automated quality inspection of painted metal parts requires robust visual anomaly detection (VAD) capable of identifying defects despite limited training data and unpredictable anomalies. Supervised algorithms are often impractical due to the limited availability of labeled anomaly data, yet existing unsupervised and semi-supervised methods continue to face accuracy challenges in real-world applications. This work addresses these challenges by introducing the KM-IAD (Konica Minolta Industrial Anomaly Detection Dataset), a new dataset comprising 2384 images categorized into 13 distinct classes of metal parts. It contains representative examples of common defects encountered in metal part manufacturing, including scratches and coating imperfections. The dataset is publicly available to facilitate further research in visual anomaly detection. We comprehensively evaluated existing state-of-the-art (SOTA) VAD methods on the KM-IAD, revealing that current approaches face significant challenges in detecting defects inherent to painted metal parts manufacturing. To address this limitation, we propose a novel self-supervised approach leveraging synthetic anomaly generation. This method achieves state-of-the-art image-level AUROC (Area Under the Receiver Operating Characteristic Curve) of 94.1% on the KM-IAD, demonstrating substantial improvements over existing techniques and providing a promising solution for high-precision industrial defect detection, ultimately reducing production costs and improving product quality.