mahalanobisad

Modeling the Distribution of Normal Data in Pre-Trained Deep Features for Anomaly Detection

Input

(Image from MVTec AD datasets https://www.mvtec.com/company/research/datasets/mvtec-ad/)

Input shape : (n, 3, 224, 224)

Output

Left to right: input, ground truth, predicted mask, segmentation result

Usage

Automatically downloads the onnx and prototxt files on the first run. It is necessary to be connected to the Internet while downloading.

In order to get the feature vector of the normal product, it is necessary to prepare the file of the normal product.
By default, normal files are got from the train directory.

To calculate the threshold you also need to prepare some test images and corresponding ground truth images. By default, test images are taken from the images directory and ground truth images are taken from the gt_masks directory.

For the sample image, first download MVTec AD datasets and place bottle/train/good/*.png files to the train directory.

For the sample image.

$ python3 mahalanobisad.py

If you want to specify the input image, put the image path after the --input option.
You can use --savepath option to change the name of the output file to save.

$ python3 mahalanobisad.py --input IMAGE_PATH --savepath SAVE_IMAGE_PATH

The feature vectors created from files in the train directory are saved to the pickle file.
From the second time, by specifying the pickle file by --feat option, it can omit the calculation of the feature vector of the normal product.
The name of the pickle file created is the same as the name of a normal product file directory.

$ python3 mahalanobisad.py --feat train.pkl

You can specify the directory of normal product files with the --train_dir option.

$ python3 mahalanobisad.py --train_dir train

The threshold is calculated using the test images and the ground truth images, but you can also give a threshold.

$ python3 mahalanobisad.py --threshold 300

By adding the --aug option, you can process with augmentation.
(default is processing without augmentation)

$ python3 mahalanobisad.py --aug

By adding the --aug_num option, you can specify amplification factor of data by augmentation. (default is 5)

$ python3 mahalanobisad.py --aug --aug_num 5

Reference

MahalanobisAD-pytorch

Framework

Pytorch

Model Format

ONNX opset=11

Netron

efficientnet-b0.onnx.prototxt
efficientnet-b1.onnx.prototxt
efficientnet-b2.onnx.prototxt
efficientnet-b3.onnx.prototxt
efficientnet-b4.onnx.prototxt
efficientnet-b5.onnx.prototxt
efficientnet-b6.onnx.prototxt
efficientnet-b7.onnx.prototxt