MvtecAD unsupervised Anomaly Detection

This respository is the unofficial implementations of DFR: Deep Feature Reconstruction for Unsupervised Anomaly Segmentation

Result of 500 epochs trained model

Selects latent sizes of Autoencoder by PCA

Classes latent size Segmentation AUC Detection AUC
bottle 116 97.2771% 99.8413%
cable 557 95.5101% 84.8951%
capsule 162 98.8928% 97.3275%
carpet 245 97.9116% 90.5297%
grid 145 97.2484 79.5322%
hazelnut 459 98.5848% 100%
leather 325 98.8649% 95.4484%
metal_nut 380 96.127% 97.263%
pill 292 98.0543% 94.108%
screw 283 99.3001% 92.0066%
tile 557 89.4887% 91.7388%
toothbrush 243 98.6729% 91.3889%
transistor 333 83.9157% 89.0833%
wood 364 91.7027% 98.9474%
zipper 115 95.6663% 83.2983%
mean 95.8141% 92.3606

How to run

requirements

pytorch scikit-learn matplotlib numpy pandas PIL wget

Train

python main.py --mode          train      
               --data_dir_head [Datapath] 
               --BATCH_SIZE    [BATCH_SIZE] 
               --LR            [Learning Rate] 
               --EPOCH         [Epochs] 
               --backbone      [Feature map of Conv in VGG19]
               --latent_dim    [Latent size of CAE] 
               --classes       [Default is all]

Download 500 Epochs Finetuned Models

Here provide the model of each classes in Drophox

python main.py --mode download

Evaluate the ROC-AUC of Test Set

python main.py --mode        evaluation    
               --classes     [Default is all]

Inference the model

python main.py --mode           inference    
               --heatmap_path   [Input path] 
               --heatmap_item   [Class of input] 
               --heatmap_gt     [GT path Default is None]
               --device         [cpu or cuda]
               --device         [Output path ]

Example run in main.py

if __name__ == "__main__":  
    cfg = config()
    cfg.mode = "inference"
    cfg.heatmap_path = 'mvtecad_unsupervise/bottle/test/broken_small/001.png'
    cfg.heatmap_item = 'bottle'
    cfg.heatmap_gt = 'mvtecad_unsupervise/bottle/ground_truth/broken_small/001_mask.png'
    cfg.device = 'cpu'
    cfg.heatmap_export = 'validate/Inferece.png'

validate/Inferece.png is

Code Reference

https://github.com/YoungGod/DFR

https://www.kaggle.com/danieldelro/unsupervised-anomaly-segmentation-of-screw-images

GitHub

View Github