Abstract:Current unsupervised anomaly detection approaches perform well on public datasets but struggle with specific anomaly types due to the domain gap between pre-trained feature extractors and target-specific domains. To tackle this issue, this paper presents a two-stage training strategy, called \textbf{ToCoAD}. In the first stage, a discriminative network is trained by using synthetic anomalies in a self-supervised learning manner. This network is then utilized in the second stage to provide a negative feature guide, aiding in the training of the feature extractor through bootstrap contrastive learning. This approach enables the model to progressively learn the distribution of anomalies specific to industrial datasets, effectively enhancing its generalizability to various types of anomalies. Extensive experiments are conducted to demonstrate the effectiveness of our proposed two-stage training strategy, and our model produces competitive performance, achieving pixel-level AUROC scores of 98.21\%, 98.43\% and 97.70\% on MVTec AD, VisA and BTAD respectively.

What problem does this paper attempt to address?

The paper primarily addresses a key issue in industrial anomaly detection: how to train a model that can effectively detect and locate various types of anomalies when only normal samples are available. Specifically, current unsupervised anomaly detection methods perform well on public datasets but face challenges in detecting specific types of anomalies, mainly due to the domain gap between pre-trained feature extractors and the target-specific domain (e.g., industrial images). To overcome this challenge, the paper proposes a two-stage training strategy called ToCoAD (Two-Stage Contrastive Learning for Industrial Anomaly Detection). The core ideas of this strategy are as follows: 1. **First Stage**: Train a discriminative network by synthesizing anomalies. This stage uses self-supervised learning by injecting synthetic anomalies (e.g., using Perlin noise) into normal images to train a discriminative network to roughly locate these anomalous regions. 2. **Second Stage**: Fine-tune the feature extractor using guided contrastive learning. In this stage, the discriminative network trained in the first stage is used to provide negative sample guidance, assisting the contrastive learning network in further optimizing the feature extractor. In this way, the model can gradually learn the anomaly distribution specific to the industrial dataset, thereby improving its generalization ability to different types of anomalies. Through this two-stage approach, ToCoAD can effectively train a model capable of accurately identifying multiple types of anomalies without any anomaly samples. Experimental results show that ToCoAD achieves highly competitive performance on several popular datasets, particularly in pixel-level anomaly detection, with its AUROC scores significantly higher than existing methods. In summary, the main contribution of this research is the proposal of a novel two-stage training strategy aimed at addressing the domain gap issue in industrial anomaly detection and improving the model's generalization performance through contrastive learning.

ToCoAD: Two-Stage Contrastive Learning for Industrial Anomaly Detection

Learning Discrimination from Contaminated Data: Multi-Instance Learning for Unsupervised Anomaly Detection

Learning Feature Inversion for Multi-class Anomaly Detection under General-purpose COCO-AD Benchmark

UAC-AD: Unsupervised Adversarial Contrastive Learning for Anomaly Detection on Multi-modal Data in Microservice Systems

A two-stage adversarial Transformer based approach for multivariate industrial time series anomaly detection

Self-Supervised Learning for Industrial Image Anomaly Detection by Simulating Anomalous Samples

Advancing Pre-trained Teacher: Towards Robust Feature Discrepancy for Anomaly Detection

Focus Your Distribution: Coarse-to-Fine Non-Contrastive Learning for Anomaly Detection and Localization

Unilaterally Aggregated Contrastive Learning with Hierarchical Augmentation for Anomaly Detection

AnomalyNCD: Towards Novel Anomaly Class Discovery in Industrial Scenarios

DTAAD: Dual Tcn-Attention Networks for Anomaly Detection in Multivariate Time Series Data

COFT-AD: COntrastive Fine-Tuning for Few-Shot Anomaly Detection

Exploring Large Vision-Language Models for Robust and Efficient Industrial Anomaly Detection

GeneralAD: Anomaly Detection Across Domains by Attending to Distorted Features

GCCAD: Graph Contrastive Learning for Anomaly Detection

GCCAD: Graph Contrastive Coding for Anomaly Detection

Unraveling the "Anomaly" in Time Series Anomaly Detection: A Self-supervised Tri-domain Solution

Abnormal Event Detection Using Deep Contrastive Learning for Intelligent Video Surveillance System

Towards Continual Adaptation in Industrial Anomaly Detection