Anomaly Detection for Real-World Industry Manufactured Products Using Computer Vision
Pawan Kumar R, M S Shreyas, Mohana · 2024
Human cognition naturally excels at identifying irregular patterns in images, enabling the distinction between expected variations and anomalies. Industrial defect classification presents unique challenges, encompassing a wide range of potential errors, from minor nuances to critical structural issues. This research work introduces PaDiM (Patch Distribution Modeling Framework) model, a novel approach for cold-start anomaly detection in industrial imagery. Prior research predominantly focuses on acquiring a model representing the standard distribution, often achieved through techniques like auto encoding, GANs, or other unsupervised adaptation methods. Even without explicit adaptation, these models exhibit robust anomaly detection capabilities, effectively localizing defects within the spatial context. PaDiM excels in industrial anomaly localization, positioning itself as a state-of-the-art solution. It maintains efficiency with swift inference times, obviating the need for specific dataset training. This characteristic renders PaDiM highly attractive for practical applications in industrial anomaly detection. In this research work, PaDiM model is implemented for real-world industry manufactured products such as Bottle, Cable, Capsule, Hazelnut and Screw using Computer Vision for anomaly detection on MVTec AD dataset. And obtained an accuracy of 83%. Additional experiments highlight PaDiM high sample efficiency, matching the performance of existing anomaly detection methods while utilizing only a fraction of the nominal training data. Furthermore, PaDiMs versatility extends to its adaptability in various industrial settings. Its ability to effectively identify and classify a wide range of defects, from subtle imperfections to major structural discrepancies, showcases its potential for widespread applicability across diverse manufacturing processes. This adapta bility is a testament to Patch Core's robustness and underscores its significance as a powerful tool in modern industrial quality control efforts.