Hierarchical Compression and Compensation for Anomaly Detection

Jingjing Fu

Abstract: Targeting for detecting anomalies of various sizes for complicated normal patterns, we propose a Hierarchical Compression and Compensation method, which introduces two key techniques, scalable bottleneck compression and template-guided compensation, for anomaly-free feature restoration. Specially, our framework compresses image features by scalable bottleneck to preserve the most crucial features shared among normal samples, so that anomalous features could be filtered out during inference. Since the image features are distorted after compression, we choose the most similar normal sample as the template, and leverage the hierarchical features from the template to compensate the distorted features for anomaly-free feature restoration. Experimental results demonstrate the effectiveness of our approach, which achieves the state-of-the-art performance on the MVTec LOCO AD dataset.