quix12345 / PEN-Net-Keras-Img_Inpainting

Image Inpainting PEN-Net CVPR-2019 Keras

PEN-Net Keras-Master

Image Inpainting task

Introduction

PEN-Net has been written into Keras now! which means it is easier to use !

"PEN-Net" is CVPR-2019 paper for Image Inpainting task, see reference for more details!

Demo

PEN-Net Decoding process visualization:

Demos:

Notice: This model havs some unknown issue which need to be fixed as PEN-Net(Keras) can not produce high quality visual realistic details, I am trying to fix it currently. If you know what's the problem, feel free to do the pull request!

Requirements

• scipy==1.1.0

• pillow

• numpy

• tensorflow-gpu

• keras

• matplotlib

• (programed in python 3.6)

Test

1. put correct model under the "/models" folder
2. init the PENNet class and call the class function "test_app_console" for testing
3. The purely white pix of the input img will be classify as the "mask" automatically
4. the output image will be in the "/generated_Imgs/test_app"

Train

1. init the PENNet class and the relevant configs (batch_size,dataset_path,etc.)
2. select a proper type(or way) of how dataloader works
3. just do the training by running the program! The model and epcho of saving will be saved directly by PENNet class automatically

Reference

1. Arxiv Paper
2. Pytorch-version

License

Licensed under an MIT license.