利用模擬異常進行自監督影像異常檢測及定位

Abstract

在圖像異常檢測的領域，異常的部分通常是少見而且無法預測的。因此，我們的目標是構建一個檢測架構，能夠在只有正常數據的情況下檢測未知的異常。在本文中，我們介紹了一個兩階段架構，其使用自監督學習來檢測和定位圖像中的異常。我們利用設計的數據增強策略來模擬真實的異常，讓模型學習區分正常數據和合成的異常數據。此外，我們比較了兩種可以結合不同層級語意特徵的方法，這兩種方法在異常檢測上都獲得了不錯的結果。無需額外的訓練資料和預訓練模型，本文提出的方法在MVTec AD 基準數據集上達到了96.4% AUROC的異常檢測分數及96.1% AUROC的異常定位分數，足以和現有的論文方法競爭。此結果展現了我們的方法在產業應用中的潛力。

In visual anomaly detection, anomalies are often rare and unpredictable. For this reason, we aim to build a detection framework that can detect unseen anomalies with only anomaly-free examples. In this paper, we introduce a two-stage framework for detecting and localizing anomalies in images using self-supervised learning. We simulate anomalies through the designed augmentation strategies, and the model learns to distinguish normal data from synthetic anomalies. In addition, we compare two methods for combining representations from different semantic levels of our network, and both of the methods obtain competitive results for defect detection. Without extra training samples and pre-trained models, the proposed approach achieves 96.4% detection AUROC and 96.1% localization AUROC on the MVTec AD benchmark, which is competitive against existing unsupervised methods. The results demonstrate the potential of our method for industrial applications.