This project implements an end-to-end anomaly detection pipeline for industrial surface inspection using the MVTec Anomaly Detection (AD) dataset. It combines classical computer vision, deep learning autoencoders, and feature-embedding methods like PaDiM to detect subtle defects across multiple object categories.
Structured loaders for all 15 object and texture categories.
Traditional image processing techniques for quick, interpretable benchmarks.
Pixel-level reconstruction error for detecting irregularities.
Embedding extraction from pretrained backbones and multivariate Gaussian modeling for high-quality anomaly maps.
Supports ROC-AUC, pixel-AUC, and visualization of anomaly heatmaps.
Easy to extend with new models or feature extractors.