With a generative adversarial network (CNN U-net framework), which can generate new defect images from normal images. This project is a simple implementation of defect fusion gan.

1. make data folder

   data
   ├── train
   │ ├── image0.png
   │ └── image1.png
   | └── ...
   └── trainannot
   │ ├── image0.png
   │ └── image1.png
   | └── ...
   

2. annot are mask of defects, which are 0 for normal pixels and non-0 for defect pixels

• run python train.py to train the model

• run python test.py to test the model, here provides a pretrained weights dftg.w with my own private dataset.

This is a MvTec test demo without training: defect defect mask mask target target target target result

Recommended images in grayscale.