[Paper Review] Unsupervised Two-Stage Anomaly Detection

Yunfei Liu, Chaoqun Zhuang+1 more|arXiv (Cornell University)|Mar 22, 2021

Anomaly Detection Techniques and Applications33 references18 citations

TL;DR

This paper proposes Unsupervised Two-stage Anomaly Detection (UTAD), a novel framework that decouples anomaly-free reconstruction from high-fidelity detail restoration using two dedicated networks: an Impression Extractor (IE-Net) and an Expert-Net. By generating an anomaly-free 'impression' first and then refining it with high-fidelity details, UTAD avoids the trade-off between coarse reconstruction artifacts and anomaly leakage, achieving state-of-the-art performance on four real-world anomaly detection datasets.

ABSTRACT

Anomaly detection from a single image is challenging since anomaly data is always rare and can be with highly unpredictable types. With only anomaly-free data available, most existing methods train an AutoEncoder to reconstruct the input image and find the difference between the input and output to identify the anomalous region. However, such methods face a potential problem - a coarse reconstruction generates extra image differences while a high-fidelity one may draw in the anomaly. In this paper, we solve this contradiction by proposing a two-stage approach, which generates high-fidelity yet anomaly-free reconstructions. Our Unsupervised Two-stage Anomaly Detection (UTAD) relies on two technical components, namely the Impression Extractor (IE-Net) and the Expert-Net. The IE-Net and Expert-Net accomplish the two-stage anomaly-free image reconstruction task while they also generate intuitive intermediate results, making the whole UTAD interpretable. Extensive experiments show that our method outperforms state-of-the-arts on four anomaly detection datasets with different types of real-world objects and textures.

Motivation & Objective

• To address the inherent contradiction in unsupervised anomaly detection where coarse reconstructions introduce false anomalies and high-fidelity reconstructions may include anomalous regions.
• To develop a framework that enables both anomaly-free and high-fidelity image reconstruction without requiring any anomaly data during training.
• To improve interpretability in anomaly detection by introducing intermediate, intuitive reconstructions (the 'impression').
• To achieve state-of-the-art performance on diverse real-world datasets with varying object types and textures.

Proposed method

• The Impression Extractor (IE-Net) generates an anomaly-free reconstruction (the 'impression' m) from the input image x, using a reconstruction loss and a mutual information maximization objective to ensure structural fidelity without anomalies.
• The Expert-Net performs high-fidelity detail restoration on the impression m to produce a final reconstructed image x̂, using an invertible mapping and a perceptual loss to preserve fine details.
• A Perceptual Measurement (PM) is employed to compute the anomaly map e by comparing differences across x, m, x̂, and m̂ using features from multiple VGG-19 layers.
• The framework uses a two-stage training process: first training IE-Net to extract the anomaly-free impression, then training Expert-Net to refine it with high-fidelity details.
• The method incorporates a Detail Guidance module (E_S) that uses skip connections from the input to guide feature adaptation in the Expert-Net, improving alignment and reconstruction quality.
• The model is trained solely on normal (anomaly-free) data, with no need for anomaly annotations or data augmentation.

Experimental results

Research questions

• RQ1 Can a two-stage reconstruction framework effectively decouple anomaly-free structure learning from high-fidelity detail restoration in unsupervised anomaly detection?
• RQ2 Does the introduction of an intermediate 'impression' m as an anomaly-free reconstruction improve detection performance and interpretability?
• RQ3 How does the proposed Perceptual Measurement (PM) compare to pixel-wise or single-layer feature differences in detecting anomaly-related discrepancies?
• RQ4 To what extent does model capacity affect anomaly detection performance, and does the two-stage design mitigate the negative impact of increased parameters?
• RQ5 How robust is the method across diverse real-world datasets with varying textures and object categories?

Key findings

• UTAD achieves state-of-the-art performance on four benchmark datasets (Bottle, Cable, Capsule, Hazelnut) from the MVTec AD dataset, with an AuROC of 0.976 on Hazelnut, significantly outperforming prior methods.
• The Perceptual Measurement (PM), combining features from 'conv1_2', 'conv2_2', and 'conv3_4' in VGG-19, yields the best anomaly detection performance, with an AuROC of 0.65 on Hazelnut, surpassing other measurement combinations.
• Ablation studies show that removing the mutual information loss (Without L_info) leads to incorrect anomaly localization, demonstrating its importance in clean impression extraction.
• Removing the Expert-Net (Without EN) results in misaligned edges and false positives, especially on fine structures like capsules, proving its critical role in detail restoration.
• The removal of the Detail Guidance module (Without E_S) causes anomaly-free regions to be falsely detected as anomalous, confirming its value in preserving spatial consistency.
• Despite higher parameter counts, standard AutoEncoders (AE-256) perform worse than UTAD, confirming that increased capacity alone does not improve detection and that the two-stage design is essential for performance.