# Generative Image Inpainting with Contextual Attention

**Authors:** Jiahui Yu, Zhe Lin, Jimei Yang, Xiaohui Shen, Xin Lu, Thomas S. Huang  
**arXiv:** [1801.07892](https://arxiv.org/abs/1801.07892)  
**Date:** 2018-01-24  
**Summary page:** https://www.summarizepaper.com/en/paper/1801.07892v1/

## Key points

- Recent deep learning approaches have shown promising results on image inpainting
- Traditional texture and patch synthesis approaches are better suited for borrowing textures from surrounding regions and utilizing surrounding image features as references during network training
- This paper presents a novel deep learning-based approach that incorporates contextual attention and borrowing information from surrounding regions
- The model is a feed-forward, fully convolutional neural network that can process images with multiple holes at arbitrary locations and variable sizes during test time
- Experiments were conducted on multiple datasets including faces, textures, and natural images to evaluate the proposed approach
- Results demonstrate that the method generates higher-quality inpainting results compared to existing approaches
- The authors will release the code and trained models for further research and application

## Summary

Recent deep learning based approaches have shown promising results on image inpainting, which involves filling in large missing regions in an image. Traditional texture and patch synthesis approaches are better suited for borrowing textures from surrounding regions and explicitly utilizing surrounding image features as references during network training to make better predictions. To address the limitations of previous methods, this paper presents a novel deep learning-based approach that incorporates contextual attention and borrowing information from surrounding regions. The model is a feed-forward, fully convolutional neural network that can process images with multiple holes at arbitrary locations and variable sizes during test time. The authors conduct experiments on multiple datasets including faces, textures, and natural images to evaluate the proposed approach. The results demonstrate that their method generates higher-quality inpainting results compared to existing approaches. They also mention that they will release the code and trained models for further research and application. In summary, this paper presents an effective approach for image inpainting that produces visually plausible and high-quality results across various datasets.