GAN collaborative image inpainting

Collaborative Sampling for Image Inpainting

Author

• Thevie Mortiniera

Inpainting on FASHION-MNIST

Visual Results

Sampling results for image inpainting by targeting the corrupted region. (Top) Input data with masked region (second row) Semantic Image Inpainting (third row) Heatmap highlighting visual differences between the inpainted results in the 2nd row and the refined results in the fourth row. The closer to the red, the higher the differences (fourth row) Collaborative Image Inpainting (bottom) Original images.

Quantitative Results :

PSNR scores, from left to right in the images above :

Method Img1 Img2 Img3 Img4 Img5 Img6 Img7 Img8 Img9 Img10 Img11 Img12 Semantic Image Inpainting 13.31 21.07 25.54 29.93 28.39 28.19 28.94 25.25 27.07 34.80 20.07 34.63 Collaborative Image Inpainting 14.65 23.84 28.63 23.43 24.53 26.77 29.22 26.57 28.18 38.27 20.10 35.97

Average scores on a test set of 2000 images :

Method SSIM PSNR IS Semantic Image Inpainting 0.813 23.713 4.160 ± 0.118 Collaborative Image Inpainting 0.834 24.478 4.184 ± 0.192

Documentation

Download dataset

The following command allow to download the FASHION-MNIST data set and create the corresponding folders as in the directory hierarchy below.

python download.py fashion_mnist

Directory hierarchy

If using an already pretrained DCGAN model, its root folder should be placed at the same hierarchy level as the collaborative-image-inpainting and Data folders, e.g below, with a pretrained model from fashion_mnist.

.
│   collaborative-image-inpainting
│   ├── src
│   │   ├── collaborator.py
│   │   ├── dataset.py
│   │   ├── dcgan.py
│   │   ├── download.py
│   │   ├── inpaint_main.py
│   │   ├── inpaint_model.py
│   │   ├── inpaint_solver.py
│   │   ├── main.py
│   │   ├── mask_generator.py
│   │   ├── ops.py
│   │   ├── policy.py
│   │   ├── solver.py
│   │   ├── tensorflow_utils.py
│   │   └── utils.py
│   │   └── utils_2.py
│   Data
│   ├── fashion_mnist
│   │   ├── train
│   │   └── val
│   fashion_mnist
│   ├── images
│   ├── inpaint
│   ├── logs
│   ├── model
│   ├── sample
│   ├── vectors

Run the app

• First of all, one need to train a DCGAN model on the choosen dataset.
• Then, use the pretrained DCGAN model to compute, offline, the closest latent vectors encodings of the images in the training set to be used during the collaborative sampling scheme.
• Finally, use the pretrained DCGAN model along with the saved latent vectors to experiment and compare the collaborative image inpainting scheme against the previous semantic image inpainting method.

Training

As an example, use the following command to train the DCGAN model. Other arguments are available in the main.py file to use different parameters.

python main.py --is_train=true --iters=25000 --dataset=fashion_mnist

Offline computing of closest latent vectors encoding

python inpaint_main.py --offline=true --dataset=fashion_mnist

Experiment between the collaborative scheme and original inpainting method.

Two modes are available between [inpaint | standard] to choose between collaborative image inpainting and standard collaborative sampling scheme. Other arguments are available in the inpaint_main.py file to use different parameters.

python inpaint_main.py --mode=inpaint --dataset=fashion_mnist

Attribution / Thanks

• This project borrowed some readme formatting and code from ChengBinJin , mostly regarding the inpainting process.
• Most of the collaborative sampling scheme was borrowed from vita-epfl