PaDiM-EfficientNet

There are two differences from the existing PaDiM code.

1. used the transfer-learned EfficientNet model, and utilized the beginning, middle, and end of feature maps of each block (Here, it is expressed in Fst, Md, Lst)
2. The test code of the existing code takes a long time by calculating the distance of Mahalanobis based on scipy, which is replaced by a tensor (attach GPU) base to reduce the inference time.

Requirement

• pytorch=1.8.0=py3.7_cuda11.1_cudnn8.0.5_0 (Or something similar)

• EfficientNet-PyTorch

  The test equipment is a single rtx3090 and see the environment.yaml or requirements.txt file for detailed environment settings.

Datasets

MVTec AD datasets : Download from MVTec website

Usage

# training
python main.py -d mvtec_anomaly_detection/ -s results --training
# test -> exclude --training option

Results

Implementation results on MVTec

• Image-level anomaly detection accuracy (ROCAUC)

MvTec	R18-Rd100	WR50-Rd550	Effi-B7-Fst-Block	Effi-B7-Md-Block	Effi-B7-Lst-Block
Carpet	0.984	0.999	0.998	0.998	0.999
Grid	0.898	0.957	0.992	0.987	0.988
Leather	0.988	1.0	1.0	1.0	1.0
Tile	0.959	0.974	0.983	0.989	0.988
Wood	0.990	0.988	0.987	0.989	0.990
All texture classes	0.964	0.984	0.992	0.992	0.993
Bottle	0.996	0.998	1.0	0.999	0.999
Cable	0.855	0.922	0.951	0.955	0.961
Capsule	0.870	0.915	0.924	0.921	0.915
Hazelnut	0.841	0.933	0.809	0.826	0.835
Metal nut	0.974	0.992	0.997	0.995	0.995
Pill	0.869	0.944	0.979	0.984	0.980
Screw	0.745	0.844	0.865	0.911	0.926
Toothbrush	0.947	0.972	0.967	0.964	0.944
Transistor	0.925	0.978	0.995	0.999	0.999
Zipper	0.741	0.909	0.930	0.927	0.922
All object classes	0.876	0.941	0.941	0.948	0.947
All classes	0.905	0.955	0.958	0.963	0.963

• Pixel-level anomaly detection accuracy (ROCAUC)

MvTec	R18-Rd100	WR50-Rd550	Effi-B7-Fst-Block	Effi-B7-Md-Block	Effi-B7-Lst-Block
Carpet	0.988	0.990	0.988	0.985	0.985
Grid	0.936	0.965	0.979	0.965	0.963
Leather	0.990	0.989	0.991	0.987	0.987
Tile	0.917	0.939	0.943	0.929	0.926
Wood	0.940	0.941	0.944	0.936	0.934
All texture classes	0.953	0.965	0.969	0.960	0.959
Bottle	0.981	0.982	0.983	0.980	0.978
Cable	0.949	0.968	0.976	0.978	0.978
Capsule	0.982	0.986	0.986	0.986	0.986
Hazelnut	0.979	0.979	0.972	0.969	0.970
Metal nut	0.967	0.971	0.962	0.962	0.959
Pill	0.946	0.961	0.964	0.961	0.954
Screw	0.972	0.983	0.985	0.987	0.987
Toothbrush	0.986	0.987	0.989	0.989	0.988
Transistor	0.968	0.975	0.971	0.981	0.982
Zipper	0.976	0.984	0.976	0.972	0.971
All object classes	0.971	0.978	0.976	0.977	0.975
All classes	0.965	0.973	0.974	0.971	0.970

• Test inference time(s)

MvTec	R18-Rd100 (original code)	WR50-Rd550 (original code)	Effi-B7	Effi-B4
Carpet	4.227	27.178	0.744	0.547
Grid	3.012	23.056	0.672	0.462
Leather	4.495	27.575	0.755	0.540
Tile	4.233	26.915	0.743	0.551
Wood	3.038	24.067	0.672	0.464
Bottle	3.169	24.119	0.669	0.469
Cable	5.332	28.246	0.786	0.596
Capsule	4.830	26.662	0.755	0.550
Hazelnut	4.033	25.470	0.730	0.540
Metal nut	4.181	25.869	0.744	0.546
Pill	5.842	30.071	0.825	0.546
Screw	5.765	29.405	0.818	0.578
Toothbrush	1.764	20.135	0.590	0.387
Transistor	3.701	24.287	0.715	0.484
Zipper	5.337	28.614	0.837	0.587

ROC Curve

• ResNet18

• Wide_ResNet50_2

• Efficientnet-b7_fst

• Efficientnet-b7_md

• Efficientnet-b7_lst

Localization examples

• sampled by EfficientNet-B7

Reference

[1] Thomas Defard, Aleksandr Setkov, Angelique Loesch, Romaric Audigier. PaDiM: a Patch Distribution Modeling Framework for Anomaly Detection and Localization. https://arxiv.org/pdf/2011.08785

[2] https://github.com/xiahaifeng1995/PaDiM-Anomaly-Detection-Localization-master