DSR -- A dual subspace re-projection network for surface anomaly detection

TL;DR

The paper introduces DSR, a novel unsupervised surface anomaly detection method that generates anomalies at the feature level, avoiding the need for synthetic image-level data, and achieves state-of-the-art results on multiple datasets.

Contribution

DSR is the first to generate near-in-distribution anomalies at the feature level using dual decoders, eliminating reliance on synthetic image augmentation.

Findings

DSR outperforms previous methods by 10% AP in anomaly detection.

DSR improves anomaly localization by 35% AP.

Achieves state-of-the-art results on KSDD2 and MVTec datasets.

Abstract

The state-of-the-art in discriminative unsupervised surface anomaly detection relies on external datasets for synthesizing anomaly-augmented training images. Such approaches are prone to failure on near-in-distribution anomalies since these are difficult to be synthesized realistically due to their similarity to anomaly-free regions. We propose an architecture based on quantized feature space representation with dual decoders, DSR, that avoids the image-level anomaly synthesis requirement. Without making any assumptions about the visual properties of anomalies, DSR generates the anomalies at the feature level by sampling the learned quantized feature space, which allows a controlled generation of near-in-distribution anomalies. DSR achieves state-of-the-art results on the KSDD2 and MVTec anomaly detection datasets. The experiments on the challenging real-world KSDD2 dataset show that DSR significantly outperforms other unsupervised surface anomaly detection methods, improving the previous top-performing methods by 10% AP in anomaly detection and 35% AP in anomaly localization.