论文原文

简述

PaDiM is a patch based algorithm. It relies on a pre-trained CNN feature extractor. The image is broken into patches and embeddings are extracted from each patch using different layers of the feature extractors. The activation vectors from different layers are concatenated to get embedding vectors carrying information from different semantic levels and resolutions. This helps encode fine grained and global contexts. However, since the generated embedding vectors may carry redundant information, dimensions are reduced using random selection. A multivariate gaussian distribution is generated for each patch embedding across the entire training batch. Thus, for each patch of the set of training images, we have a different multivariate gaussian distribution. These gaussian distributions are represented as a matrix of gaussian parameters.

During inference, Mahalanobis distance is used to score each patch position of the test image. It uses the inverse of the covariance matrix calculated for the patch during training. The matrix of Mahalanobis distances forms the anomaly map with higher scores indicating anomalous regions.

Architecture

Patch Distribution Modeling

特征提取

正样本学习

计算异常图

补充知识

PRO-score

马氏距离

SPADE

Sub-Image Anomaly Detection with Deep Pyramid Correspondences

Nearest neighbor (kNN) methods utilizing deep pre-trained features exhibit very strong anomaly detection performance when applied to entire images. A limitation of kNN methods is the lack of...

https://arxiv.org/abs/2005.02357

实验

消融实验

未对齐数据集

可扩展性

参考

Padim - Anomalib v0.3.7

PaDiM is a patch based algorithm. It relies on a pre-trained CNN feature extractor. The image is broken into patches and embeddings are extracted from each patch using different layers of the feature extractors. The activation vectors from different layers are concatenated to get embedding vectors carrying information from different semantic levels and resolutions. This helps encode fine grained and global contexts. However, since the generated embedding vectors may carry redundant information, dimensions are reduced using random selection. A multivariate gaussian distribution is generated for each patch embedding across the entire training batch. Thus, for each patch of the set of training images, we have a different multivariate gaussian distribution. These gaussian distributions are represented as a matrix of gaussian parameters.

https://anomalib.readthedocs.io/en/v0.3.7/reference_guide/algorithms/padim.html

PaDiM: a Patch Distribution Modeling Framework for Anomaly...

We present a new framework for Patch Distribution Modeling, PaDiM, to concurrently detect and localize anomalies in images in a one-class learning setting. PaDiM makes use of a pretrained...

https://arxiv.org/abs/2011.08785

Sub-Image Anomaly Detection with Deep Pyramid Correspondences

Nearest neighbor (kNN) methods utilizing deep pre-trained features exhibit very strong anomaly detection performance when applied to entire images. A limitation of kNN methods is the lack of...

https://arxiv.org/abs/2005.02357