F2PAD: A General Optimization Framework for Feature-Level to Pixel-Level Anomaly Detection

TL;DR

F2PAD introduces a universal optimization framework that improves pixel-level anomaly detection accuracy in image inspection systems by refining feature-based methods, especially in boundary localization, using feature-level guidance during inference.

Contribution

The paper presents a novel, plug-and-play optimization framework that enhances existing feature-based anomaly detection methods with limited assumptions, improving boundary accuracy.

Findings

Enhanced boundary localization in anomaly detection

Effective when integrated with popular backbone methods

Universal applicability across different feature-based models

Abstract

Image-based inspection systems have been widely deployed in manufacturing production lines. Due to the scarcity of defective samples, unsupervised anomaly detection that only leverages normal samples during training to detect various defects is popular. Existing feature-based methods, utilizing deep features from pretrained neural networks, show their impressive performance in anomaly localization and the low demand for the sample size for training. However, the detected anomalous regions of these methods always exhibit inaccurate boundaries, which impedes the downstream tasks. This deficiency is caused: (i) The decreased resolution of high-level features compared with the original image, and (ii) The mixture of adjacent normal and anomalous pixels during feature extraction. To address them, we propose a novel unified optimization framework (F2PAD) that leverages the Feature-level information to guide the optimization process for Pixel-level Anomaly Detection in the inference stage. The proposed framework is universal and plug-and-play, which can enhance various feature-based methods with limited assumptions. Case studies are provided to demonstrate the effectiveness of our strategy, particularly when applied to three popular backbone methods: PaDiM, CFLOW-AD, and PatchCore.