This repository contains a Jupyter Notebook that demonstrates an anomaly detection model available in anomalib: PaDiM A Patch Distribution Modeling Framework for Anomaly Detection in MVTec dataset. This method is implemented using the anomalib library in Python.

Image Anomaly Detection

Detecting anomalies in images is a popular application of anomaly detection. In this tutorial, we will go over a popular dataset known as the "MVTec Anomaly Detection" dataset. MVTec is a dataset for benchmarking anomaly detection methods with a focus on industrial inspection.

Here are some features of the dataset:

• Contains over 5000 high-resolution images.
• Images are divided into fifteen distinct object and texture categories such as bottle, cable, carpet, wood, leather and pill.
• Each category consists of two sets: training images (defect-free) and test images (with defects and without defects).

Some example of the MVTec dataset:

Anomalib:

Anomalib is a deep learning library for developing and deploying state-of-the-art anomaly detection algorithms for benchmarking on both public and private datasets.

Key features:

• The largest public collection of ready-to-use deep learning anomaly detection algorithms and benchmark datasets.
• Provides a set of tools that facilitate the development and implementation of Anomaly Detection models.
• Focus on image-based anomaly detection or anomalous pixel regions within images in a dataset.

Model

Currently, there are 13 anomaly detection models available in anomalib library. Namely,

• CFA
• PADIM
• CFlow
• DFKDE
• DFM
• DRAEM
• EfficientAd
• FastFlow
• GANomaly
• PatchCore
• Reverse Distillation
• STFPM

In this demo, we'll be using Padim.

PaDiM: A Patch Distribution Modeling Framework for Anomaly Detection

• Patch-based algorithm relying on pre-trained CNN feature extractor.
• Breaks image into patches, extracts embeddings using different layers.
• Concatenates activation vectors for diverse semantic levels.
• Encodes fine-grained and global contexts in embeddings.
• Reduces dimensions of embedding vectors to mitigate redundancy.
• Generates multivariate Gaussian distribution for each patch embedding.
• Distribution calculated across entire training batch.
• Inference uses Mahalanobis distance to score test image patches.
• Uses inverse of covariance matrix from training for Mahalanobis distance.
• Anomaly map formed from Mahalanobis distance scores.
• Higher scores in anomaly map indicate anomalous regions.