image-anomaly-det: Image Anomaly Detection

This repository provides implementation of image anomaly detection methods using PyTorch.

Implemented methods

• anotwin: AnomalyTwin, deep metric learning.
• efficient_gan: EfficientGAN, deep image reconstruction.
• dadgt: Deep Anomaly Detection Using Geometric Transformations (DADGT), kind of deep metric learning.
• img_hash: Image hashing, kind of metric learning.
• to be continued.

Setup

Download/Install

• Grad-CAM: Using pytorch_cnn_visualizations implementation.
• ArcFace: Using arcface_pytorch implementation.
• Image Hashing: Using module provided as imagehash: https://github.com/JohannesBuchner/imagehash

pip install -r requirements.txt
git clone https://github.com/daisukelab/pytorch_cnn_visualizations
git clone https://github.com/daisukelab/arcface_pytorch.git

Example Results

target	EfficientGAN	DADGT	phash=16	phash=32	phash=64	whash=16	whash=32
bottle	0.740476	0.897619	0.894048	0.769048	0.834921	0.831746	0.835714
cable	0.579648	0.885495	0.763681	0.71786	0.655454	0.688999	0.756747
capsule	0.552852	0.688073	0.732948	0.764061	0.751296	0.540088	0.713801
carpet	0.702247	0.455859	0.549559	0.476124	0.49378	0.60012	0.461276
grid	0.728488	0.523810	0.913952	0.839181	0.786967	0.619883	0.56391
hazelnut	0.597143	0.600714	0.78625	0.83625	0.8125	0.462679	0.530714
leather	0.503736	0.616168	0.946671	0.897079	0.757812	0.838655	0.793648
metal_nut	0.537146	0.867546	0.660557	0.597996	0.556452	0.714809	0.782014
pill	0.696672	0.680715	0.649345	0.628205	0.635706	0.648663	0.854883
screw	0.322402	0.573683	0.598176	0.686719	0.647469	0.573786	0.605759
tile	0.481602	0.501804	0.440657	0.478896	0.450397	0.330447	0.308983
toothbrush	0.586111	0.922222	0.954167	0.969444	0.915278	0.713889	0.806944
transistor	0.53875	0.904583	0.898125	0.872708	0.814375	0.723125	0.743958
wood	0.892982	0.849123	0.502632	0.550439	0.710088	0.393421	0.321053
zipper	0.52521	0.870273	0.737526	0.794118	0.874081	0.463629	0.419118