Image Anomaly Detection With Semantic- Enhanced Augmentation and Distributional Kernel

In the field of image anomaly detection, self-supervised learning using only normal samples has become a hot research topic since it is difficult to acquire sufficient abnormal data. However, existing self-supervised methods usually consider the impact of the superficial information rather than the semantics of images, resulting in limited generalisation performance. Furthermore, the small amount of training data further limits the performance of anomaly detection. To overcome both issues, we propose a two-stage self-supervised learning framework based on CutPaste with a feature extraction stage followed by an anomaly detection stage. We first add a diminished semantics augmentation to generate harder abnormal samples, so that the feature extraction model pays more attention to semantic information. Furthermore, we increase the density of normal samples with normal sample augmentations and integrate a data-dependent distributional kernel to improve the detection of abnormal images in the second stage, based on the adaptive distributional similarity. Our extensive experiments on the MVTec anomaly detection dataset demonstrate the superior performance of the proposed method.

I. Introduction

Anomaly detection is the process of detecting instances that deviate significantly from the majority of other instances in a dataset [1]. Anomaly detection has a wide range of applications in manufacturing defect detection [2], [3], road traffic monitoring [4], medical diagnostics [5] and other fields. However, the common supervised anomaly detection [6] is infeasible in practice, because data labels are difficult to acquire, and human-labelled anomaly information is often hard to obtain [7]. To tackle these limitations, collecting anomaly-free images to train anomaly detection models has became a hot research direction.