PEN-Net for Image Inpainting

Arxiv Paper | Project | Poster | BibTex

Learning Pyramid-Context Encoder Network for High-Quality Image Inpainting
Yanhong Zeng, Jianlong Fu, Hongyang Chao, and Baining Guo.
In CVPR 2019.

Introduction

Existing inpainting works either fill missing regions by copying fine-grained image patches or generating semantically reasonable patches (by CNN) from region context, while neglect the fact that both visual and semantic plausibility are highly-demanded.

Our proposals combine these two mechanisms by,

1. Cross-Layer Attention Transfer (ATN). We use the learned region affinity from high-lelvel feature maps to guide feature transfer in adjacent low-level layers in an encoder.
2. Pyramid Filling. We fill holes multiple times (depends on the depth of the encoder) by using ATNs from deep to shallow.

Example Results

We re-implement PEN-Net in Pytorch for faster speed, which is slightly different from the original Tensorflow version used in our paper. Each triad shows original image, masked input and our result.

Run

0. Requirements:
   • Install python3.6
   • Install pytorch (tested on Release 1.1.0)
1. Training:
   • Prepare training images filelist [our split]
   • Modify celebahq.json to set path to data, iterations, and other parameters.
   • Our codes are built upon distributed training with Pytorch.
   • Run python train.py -c [config_file] -n [model_name] -m [mask_type] -s [image_size] .
   • For example, python train.py -c configs/celebahq.json -n pennet -m square -s 256
2. Resume training:
   • Run python train.py -n pennet -m square -s 256 .
3. Testing:
   • Run python test.py -c [config_file] -n [model_name] -m [mask_type] -s [image_size] .
   • For example, python test.py -c configs/celebahq.json -n pennet -m square -s 256
4. Evaluating:
   • Run python eval.py -r [result_path]

Pretrained models

Download the models below and put it under release_model/

CELEBA-HQ | DTD | Facade | Places2

We also provide more results of central square below for your comparisons

CELEBA-HQ | DTD | Facade

TensorBoard

Visualization on TensorBoard for training is supported.

Run tensorboard --logdir release_model --port 6006 to view training progress.

Citation

If any part of our paper and code is helpful to your work, please generously cite with:

@inproceedings{yan2019PENnet,
  author = {Zeng, Yanhong and Fu, Jianlong and Chao, Hongyang and Guo, Baining},
  title = {Learning Pyramid-Context Encoder Network for High-Quality Image Inpainting},
  booktitle = {The IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
  pages={1486--1494},
  year = {2019}
}

License

Licensed under an MIT license.