CutPaste: Self-Supervised Learning for Anomaly Detection and Localization

Abstract

We aim at constructing a high performance model for defect detection that detects unknown anomalous patterns of an image without anomalous data. To this end, we propose a two-stage framework for building anomaly detectors using normal training data only. We first learn self-supervised deep representations and then build a generative one-class classifier on learned representations. We learn representations by classifying normal data from the CutPaste, a simple data augmentation strategy that cuts an image patch and pastes at a random location of a large image. Our empirical study on MVTec anomaly detection dataset demonstrates the proposed algorithm is general to be able to detect various types of real-world defects. We bring the improvement upon previous arts by 3.1 AUCs when learning representations from scratch. By transfer learning on pretrained representations on ImageNet, we achieve a new state-of-theart 96.6 AUC. Lastly, we extend the framework to learn and extract representations from patches to allow localizing defective areas without annotations during training.

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

“…The pixel-wise reconstruction error can be applied to localize anomalies (Bergmann et al 2019), and the image level anomaly score is thus determined by aggregating pixel-wise errors (Gong et al 2019). Despite the high interpretability of reconstruction and comparison, the pixel-wise difference fails to encode the global semantic meaning of images (Ren et al 2019;Li et al 2021). Also, the autoencoder sometimes generates comparable reconstruction results for the anomalous images too (Perera, Nallapati, and Xiang 2019).…”

“…Some methods are predicting rotation of images (Golan and El-Yaniv 2018; Bergman and Hoshen 2020) and contrastive learning with usual image augmentation strategies (Sohn et al 2021;Tack et al 2020). Although these methods well capture the semantic object information in images, they fail to encode the fine-grained local irregularities in anomalies (Li et al 2021). Thus, several works (DeVries and Taylor 2017; Yun et al 2019;Li et al 2021) created a set of new data augmentations that replicates the local defects in anomalies.…”

“…Although these methods well capture the semantic object information in images, they fail to encode the fine-grained local irregularities in anomalies (Li et al 2021). Thus, several works (DeVries and Taylor 2017; Yun et al 2019;Li et al 2021) created a set of new data augmentations that replicates the local defects in anomalies. The method Cutout (DeVries and Taylor 2017) randomly removes a small rectangular area from images, and CutPaste (Li et al 2021) further modifies the algorithm to cut a patch from one image and paste it on the other.…”

“…In the fine alignment stage, we apply self-supervised learning and propose a novel pretext task for learning the normal representation. The current state-of-the-art (Li et al 2021) in self-supervised anomaly detection designs augmentations that generate abnormal samples through mixing normal image patches. However, we lack sufficient prior knowledge of the real-world anomaly distributions, so the created defects cannot model the numerous real-life possibilities of anomalies.…”

FastFlow: Unsupervised Anomaly Detection and Localization via 2D Normalizing Flows

“…To the distribution estimation module, previous approaches used the non-parametric method to model the distribution of features for normal images. For example, they estimated the multidimensional Gaussian distribution (Li et al 2021;Defard et al 2020) by calculating the mean and variance for features, or used a clustering algorithm to estimate these normal features by normal clustering (Reiss et al 2021;Roth et al 2021). Recently, some works (Rudolph, Wandt, and Rosenhahn 2021;Gudovskiy, Ishizaka, and Kozuka 2021) began to use normalizing flow (Kingma and Dhariwal 2018) to estimate distribution.…”