Implemented color and grayscale missing pixel inpainting paper
The original image is masked like Figure 1, where the mask's pixel value is zero. If the image is RGB, then mask each channel independently. The task is to inpaint the pixel of 0 on the masked image.
Figure 1. grayscale 80% masked image
Figure 2. color 40% masked image
Figure 3. color 60% masked image
Figure 3. Inpainting CNN Result
Table 1. L2 Norm per missing pixel of example
| TYPE | 80% Grayscale | 40% Color | 60% Color |
|---|---|---|---|
| L2 Norm per missing pixel | 0.10093 | 0.03512 | 0.08130 |
The preprocessing.py will generate four .npy files, like train_x_color.npy, train_y_color.npy, train_x_gray.npy, train_y_gray.npy.
python train.py train_gray 1 (train gray model and will generate ./gray-model/
python train.py train_gray 0 (train color model and generate ./color-model/)
The pre-processed data can be downloaded on preprocessed data and pre-trained model could be download on model, where the dataset is randomly select from MSCOCO
python inpainting --img_dir some-of-missing-pixel-image --model_version XXXX (where XXXX is the best trained model)