Anomaly detection and segmentation in industrial images using multi-scale reverse distillation

Abstract

Anomaly detection and segmentation in industrial images are critical tasks requiring robust and precise methodologies. This paper presents the Multi-Scale Reverse Distillation (MSRD) methodology, an innovative improvement of the foundational reverse distillation approach. MSRD leverages autoencoder-based techniques integrated with information at different levels to significantly enhance reconstruction capabilities. A novel module incorporated at the decoder’s end facilitates precise sample reconstruction. The proposed loss function incorporates the reconstruction loss $L_{Recon}$, calculated using structural similarity index measure (SSIM) between the original and reconstructed images, in addition to the knowledge distillation loss $L_{KD}$. Additionally, the integration of a feature pyramid network improves the spatial coherence of anomaly maps across varying scales, enabling detailed anomaly segmentation. The MSRD method undergoes rigorous evaluation on three public datasets, demonstrating superior performance in both anomaly detection and segmentation. The results highlight MSRD’s adaptability and effectiveness in one-class learning-based applications. This study underscores MSRD’s potential as a powerful tool for industrial anomaly detection, offering significant advancements in AI-driven image analysis.