Constrained Adaptive Projection with Pretrained Features for Anomaly Detection

TL;DR

This paper introduces CAP, a novel anomaly detection framework that leverages pretrained features with a constrained adaptive projection to improve detection accuracy while avoiding pattern collapse.

Contribution

The paper proposes a new anomaly detection method combining a linear projection head, a reformed self-attention, and a specialized loss to enhance feature adaptation and prevent pattern collapse.

Findings

Achieves 96.5% AUROC on CIFAR-100

Achieves 97.0% AUROC on CIFAR-10

Achieves 89.9% AUROC on MvTec

Abstract

Anomaly detection aims to separate anomalies from normal samples, and the pretrained network is promising for anomaly detection. However, adapting the pretrained features would be confronted with the risk of pattern collapse when finetuning on one-class training data. In this paper, we propose an anomaly detection framework called constrained adaptive projection with pretrained features (CAP). Combined with pretrained features, a simple linear projection head applied on a specific input and its k most similar pretrained normal representations is designed for feature adaptation, and a reformed self-attention is leveraged to mine the inner-relationship among one-class semantic features. A loss function is proposed to avoid potential pattern collapse. Concretely, it considers the similarity between a specific data and its corresponding adaptive normal representation, and incorporates a constraint term slightly aligning pretrained and adaptive spaces. Our method achieves state-ofthe-art anomaly detection performance on semantic anomaly detection and sensory anomaly detection benchmarks including 96.5% AUROC on CIFAR- 100 dataset, 97.0% AUROC on CIFAR-10 dataset and 89.9% AUROC on MvTec dataset.