Real-Time Industrial Anomaly Detection via Sparse Reconstruction

In the field of industrial anomaly detection, the scarcity of anomalous data and labels poses significant challenges, necessitating models that can efficiently detect and localize anomalies with minimal reliance on anomalous training data. Traditional approaches often utilize outlier detection strategies on pre-trained features, which are hampered by the inclusion of redundant and irrelevant information, leading to decreased computational efficiency and diminished performance in real-time applications. Addressing these limitations, this paper introduces a novel defect detection and localization strategy that emphasizes rapid feature reconstruction. Our methodology comprises three key components: (1) a robust pre-trained feature extractor that generates descriptive image features, (2) an innovative feature dictionary developed via dictionary learning to embed features from normal images, and (3) a dynamic feature reconstructor designed for swift reconstruction of test features utilizing the dictionary. This approach enables precise anomaly identification and localization by assessing differences between original and reconstructed features. Rigorous testing on the MVTec AD dataset—a benchmark for real-world industrial anomaly detection—validates the method’s superiority, demonstrating substantial improvements in detection speed with minimal impact on accuracy. The findings suggest that this strategy holds significant promise for enhancing the efficiency and reliability of anomaly detection in industrial settings.