[UL] Abnormaly detection

UL stands for Unsupervised Learning.

In this application, we use the PatchCore algorithms to learn the good-only dataset. Then we use the K-Nearest Neighbors (KNN) algorithms to separate the newly seen good and ng data.

1. Prepare your dataset

# your dataset structure should be like this
data/
    -train/
        -good/
            -*.jpg
    -test/
        -good/
            -*.jpg
        -ng/
            -*.jpg

Notes

The number of train data should be > 10 items. If it is <10 items, please rotated the images to create more.

2. Usages

Training

Go to /projects/indad-holes

Execute this command:

python train.py --model_type spade --data_dir "<PATH_TO_TRAIN_FOLDER>"

Note

PATH_TO_TRAIN_FOLDER should be one-level above the image files, i.e. ../data/train, not ../data/train/good

The model will be saved into results folder.

Testing

Go to /projects/indad-holes

Execute this command:

python predict.py --model_type spade --data_dir "<PATH_TO_TEST_FOLDER>" --threshold 1

Note

PATH_TO_TRAIN_FOLDER should be where the image files are, i.e. ../data/train/good

Result images will be save into results folder.

3. Improve the performance

If the prediction is not good, please consider either tuning the threshold or adding more data