Reliable Anomaly Detection and Localization System: Implications on Manufacturing Industry

Abstract

Product quality inspection is a critical component of industrial manufacturing. An accurate and reliable Anomaly Detection and Localization (ADL) system for industrial product quality inspection is essential in real-world manufacturing factories. Collecting massive anomalous products is difficult because the number of anomalous products is limited and rare in a realistic manufacturing scenario. Therefore, the paper treats the ADL problem as a cold-start challenge, training the defects inspection network only using nominal (non-defective) images. Significantly, the paper aims to bridge the gap between academic research and real-world manufacturing industry applications. The paper lists issues that current state-of-the-art academic research faces when applied in real-world manufacturing settings, then a Reliable Anomaly Detection and Localization (RADL) system is developed to solve the issues. RADL is improved in three aspects. Firstly, the common image pre-processing method is modified by considering the characteristics of real-world industrial images. Secondly, a Fake Defect Feature Augmentation (FDFA) strategy to mitigate the scarcity of real-world data. Thirdly, a Hardness-aware Cross-Entropy loss (HCELoss) is adopted to enhance the stability and reliability of the system. On the public MVTec AD benchmarks, the proposed RADL outperforms previous methods with 99.53% in I-AUROC, 97.85% in P-AUROC, and 91.60% in PRO. Furthermore, RADL is evaluated under industrial manufacturing settings in two real-world datasets collected from industrial production lines. The experimental results demonstrate the superiority of the proposed strategies in a public dataset and real-world manufacturing industrial environments.