Coherent Semantic Attention for image inpainting (ICCV 2019)

Arxiv |ICCV 2019 paper| BibTex

Introduction

The existing inpainting methods often generate contents with blurry textures and distorted structures due to the discontinuity of the local pixels.From a semantic-level perspective, the local pixel discontinuity is mainly because these methods ignore the semantic relevance and feature continuity of hole regions. To handle this problem, we investigate the human behavior in repairing pictures and propose a fined deep generative model-based approach with a novel coherent semantic attention (CSA) layer, which can not only preserve contextual structure but also make more effective predictions of missing parts by modeling the semantic relevance between the holes features. Meanwhile, we further propose consistency loss and feature patch discriminator to stabilize the network training process and improve the details.

Model Architecture

CSA Architecture

Feature patch discriminator

Prerequisites

• Windows or Linux
• Python3
• Pytorch 1.0
• NVIDIA GPU + CUDA CuDNN.
• Jupyter Notebook

Installation

• Clone this repo:

  git clone https://github.com/KumapowerLIU/CSA-inpainting.git
  cd CSA-inpainting
  

• Install PyTorch and dependencies from http://pytorch.org

• Or you can clone the repo and run by conda

  conda env create -f environment.yaml 
  conda activate csa-inpainting
  jupyter notebook
  

  Open the browser and enter the URL http://localhost:8080 to enjoy jupyter notebook

Datasets

We use Places2, CelebA and Paris Street-View datasets. To train a model on the full dataset, download datasets from official websites.

Our model is trained on the irregular mask dataset provided by Liu et al. You can download publically available Irregular Mask Dataset from their website.

Model Training

• Download your own inpainting datasets.
• Open train.ipynb in Jupyter Notebook
• Modify Option Class to set data_root,mask_root,checkpoints_dir, mask_type and other parameters.
• Run train.ipynb

Model Testing

• Open test.ipynb in Jupyter Notebook
• Keep the same parameters as during training
• Run test.ipynb

License

CC 4.0 Attribution-NonCommercial International. The software is for educaitonal and academic research purpose only.

Citation

@InProceedings{Liu_2019_CSA,
    Author = {Hongyu Liu and Bin Jiang and Yi Xiao and Chao Yang},
    Title = {Coherent Semantic Attention for Image Inpainting},
    booktitle = { IEEE International Conference on Computer Vision (ICCV)},
    month = {July},
    year = {2019}
    
}

Acknowledgments

We benefit a lot from Shift-net