[論文レビュー] A Unified Model for Multi-class Anomaly Detection

Zhiyuan You, Lei Cui+5 more|arXiv (Cornell University)|Jun 8, 2022

Anomaly Detection Techniques and Applications被引用数 87

ひとこと要約

UniADは、再構成ベースの手法におけるアイデンティティショートカット問題を解決しつつ、層ごとのクエリデコード、隣接遮蔽付き注意、特徴のジッタリングを用いて、 fine-tuning なしで複数クラスにわたる異常を検出・局在する統一フレームワークを提示します。

ABSTRACT

Despite the rapid advance of unsupervised anomaly detection, existing methods require to train separate models for different objects. In this work, we present UniAD that accomplishes anomaly detection for multiple classes with a unified framework. Under such a challenging setting, popular reconstruction networks may fall into an "identical shortcut", where both normal and anomalous samples can be well recovered, and hence fail to spot outliers. To tackle this obstacle, we make three improvements. First, we revisit the formulations of fully-connected layer, convolutional layer, as well as attention layer, and confirm the important role of query embedding (i.e., within attention layer) in preventing the network from learning the shortcut. We therefore come up with a layer-wise query decoder to help model the multi-class distribution. Second, we employ a neighbor masked attention module to further avoid the information leak from the input feature to the reconstructed output feature. Third, we propose a feature jittering strategy that urges the model to recover the correct message even with noisy inputs. We evaluate our algorithm on MVTec-AD and CIFAR-10 datasets, where we surpass the state-of-the-art alternatives by a sufficiently large margin. For example, when learning a unified model for 15 categories in MVTec-AD, we surpass the second competitor on the tasks of both anomaly detection (from 88.1% to 96.5%) and anomaly localization (from 89.5% to 96.8%). Code is available at https://github.com/zhiyuanyou/UniAD.

研究の動機と目的

• Motivate and enable anomaly detection across multiple object classes with a single model rather than separate one-class models.
• Identify and mitigate the “identity shortcut” in reconstruction-based anomaly detection, especially under the unified multi-class setting.
• Develop architectural and training strategies that preserve multi-class distribution modeling while preventing easy reconstruction of anomalies.
• Demonstrate state-of-the-art performance on standard benchmarks (MVTec-AD and CIFAR-10) under a unified, many-versus-many setting.

提案手法

• Adopt a transformer-based feature reconstruction framework with a fixed backbone feature extractor.
• Introduce a Neighbor Masked Attention (NMA) module to prevent information leakage from input to output during reconstruction.
• Implement a Layer-wise Query Decoder (LQD) to leverage a learnable query embedding at each decoder layer, strengthening modeling of the multi-class distribution.
• Apply Feature Jittering (FJ) to perturb input features and cast the reconstruction task as denoising, improving robustness to noisy inputs.
• Train with a standard MSE loss between original and reconstructed features, enabling anomaly localization via reconstruction differences.
• Utilize a fixed EfficientNet-B4 backbone for multi-stage feature extraction and concatenate features across stages for reconstruction.

実験結果

リサーチクエスチョン

• RQ1 Can a single unified model effectively detect and localize anomalies across many object classes without class-specific fine-tuning?
• RQ2 How can reconstruction-based anomaly detection avoid the identity shortcut, especially in a unified multi-class setting?
• RQ3 Do architectural choices like query embeddings, masked attention, and feature jittering improve performance over standard transformers for anomaly detection?
• RQ4 What is the impact of these components on anomaly detection accuracy and localization under a unified setting?
• RQ5 How does UniAD perform relative to state-of-the-art methods on MVTec-AD and CIFAR-10 in unified versus separate settings?

主な発見

• UniAD achieves state-of-the-art anomaly detection and localization on MVTec-AD under the unified setting, with a large margin over the previous best method on both tasks.
• In a 15-category unified model for MVTec-AD, AUROC for anomaly detection improves from 88.1% (previous best) to 96.5%, and localization improves from 89.5% to 96.8%.
• A single UniAD model under the unified setting maintains high performance across categories without fine-tuning, showing minimal drop from separate-case performance.
• Layer-wise query embeddings, neighbor masked attention, and feature jittering collectively contribute to strong performance, with ablations confirming the importance of each component.
• On CIFAR-10 under the unified setting, UniAD outperforms baselines across four category combinations, achieving an average AUROC of 87.2% without auxiliary outlier exposure.