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

Abstract

Unsupervised anomaly detection and localization is crucial to the practical application when collecting and labeling sufficient anomaly data is infeasible. Most existing representation-based approaches extract normal image features with a deep convolutional neural network and characterize the corresponding distribution through non-parametric distribution estimation methods. The anomaly score is calculated by measuring the distance between the feature of the test image and the estimated distribution. However, current methods can not effectively map image features to a tractable base distribution and ignore the relationship between local and global features which are important to identify anomalies. To this end, we propose FastFlow implemented with 2D normalizing flows and use it as the probability distribution estimator. Our FastFlow can be used as a plug-in module with arbitrary deep feature extractors such as ResNet and vision transformer for unsupervised anomaly detection and localization. In training phase, FastFlow learns to transform the input visual feature into a tractable distribution and obtains the likelihood to recognize anomalies in inference phase. Extensive experimental results on the MVTec AD dataset show that FastFlow surpasses previous state-of-the-art methods in terms of accuracy and inference efficiency with various backbone networks. Our approach achieves 99.4% AUC in anomaly detection with high inference efficiency.

A Survey on Unsupervised Industrial Anomaly Detection Algorithms

“…However, in order to estimate the distribution, the original one-dimensional normalizing flow model must flatten the two-dimensional input feature into a one-dimensional vector, which destroys the inherent spatial positional relationship implied by the two-dimensional image and constrains the NF model. FastFlow [88] expands the original normalizing flow model to two-dimensional space to address the concerns mentioned above. As shown in Fig.…”

“…To our best knowledge, no review has been done for the recently emerged unsupervised methods. The article will provide a comprehensive and in-depth summary of the state-of-the-art algorithms for industrial anomaly detection, which will be divided into five categories listed as 4.1 Normalizing Flow (NF) based [85][86][87][88] [103,[110][111][112]. This comprehensive summary is expected to contribute to the implementation and practice of industrial field.…”

Benchmarking Unsupervised Anomaly Detection and Localization

“…For the feature extractor, existing approaches usually adopt the CNN based or ViT based model [7,8,22,34], such as the ResNet [15] or ViT [11] trained on the Ima-geNet. Furthermore, to learn the domain-specific semantic vectors for images, a series of methods [18,20,37] employed self-supervised learning to achieve the ImageNet pretrained model adaption.…”

“…Patch-Core [22] uses a maximally representative memory bank of normal patch-features and uses the k-nearest neighbor method to search for the closest feature for distance metric. Recently, some works [14,23,34] began to use parametric methods such as normalizing flow [10,16] to estimate the distribution of normal images. Through a trainable process that maximizes the log-likelihood of normal image features, they embed normal image features into standard normal distribution and use the probability to identify and locate anomalies.…”

“…CSFlow [24] proposes a novel fully convolutional cross-scale normalizing flow to jointly processes multiple feature maps of different scales. FastFlow [34] designs a 2D normalizing flow for anomaly detection and localization with fully convolutional networks and achieves the state of the art performance.…”