Self-supervised Augmented Patches Segmentation for Anomaly Detection

Jun Long, Yuxi Yang, Liujie Hua, Yiqi Ou

Abstract: In this paper, our goal is to detect unknown defects in high-resolution images in the absence of anomalous data. Anomaly detection is usually performed at image-level or pixel-level. Considering that pixel-level anomaly classification achieves better representation learning in a finer-grained manner, we regard data augmentation transforms as a self-supervised segmentation task from which to extract the critical and representative information from images. Due to the unpredictability of anomalies in real scenarios, we propose a novel abnormal sample simulation strategy which augmented patches are randomly pasted to original image to create a generalized anomalous pattern. Following the framework of self-supervised, segmenting augmented patches is used as a proxy task in the training phase to extract representation separating normal and abnormal patterns, thus constructing a one-class classifier with a robust decision boundary. During the inference phase, the classifier is used to perform anomaly detection on the test data, while directly determining regions of unknown defects in an end-to-end manner. Our experimental results on MVTec AD dataset and BTAD dataset demonstrate the proposed SSAPS outperforms any other self-supervised based methods in anomaly detection. Code is available at https://github.com/BadSeedX/SSAPS .