# Image Inpainting for Irregular Holes Using Partial Convolutions

**Authors:** Guilin Liu, Fitsum A. Reda, Kevin J. Shih, Ting-Chun Wang, Andrew Tao, Bryan Catanzaro  
**arXiv:** [1804.07723](https://arxiv.org/abs/1804.07723)  
**Date:** 2018-04-20  
**Summary page:** https://www.summarizepaper.com/en/paper/1804.07723v1/

## Key points

- Existing methods in image inpainting using deep learning techniques often suffer from color discrepancies and blurriness due to standard convolutional networks.
- The researchers, including Guilin Liu, Fitsum A. Reda, Kevin J. Shih, Ting-Chun Wang, Andrew Tao, and Bryan Catanzaro, proposed a novel approach called partial convolutions to address these limitations.
- Partial convolutions are used to mask and renormalize convolutions so they only consider information from valid pixels, improving the quality of inpainted images.
- The model demonstrated superior performance in handling irregularly shaped holes compared to existing methods.
- Qualitative and quantitative comparisons show significant improvements by leveraging partial convolutions for image inpainting.
- This study provides valuable insights into advancing image inpainting using partial convolutions for more visually appealing and accurate results with promising potential for future developments in the field.

## Summary

In the field of image inpainting, where missing or corrupted parts of an image are filled in using deep learning techniques, existing methods often suffer from issues such as color discrepancies and blurriness due to the use of standard convolutional networks. These networks typically rely on information from both valid pixels and substitute values in masked areas, leading to suboptimal results that require expensive post-processing to correct. To address these limitations, a group of researchers including Guilin Liu, Fitsum A. Reda, Kevin J. Shih, Ting-Chun Wang, Andrew Tao, and Bryan Catanzaro have proposed a novel approach called partial convolutions. are used in this method to mask and renormalize the convolutions so that they only consider information from valid pixels. This effectively improves the quality of inpainted images by reducing color discrepancies and blurriness. Additionally, is introduced for subsequent layers during the forward pass of the network. Their model has demonstrated superior performance compared to existing methods when dealing with irregularly shaped holes in images. Through qualitative and quantitative comparisons with other approaches, have been achieved by leveraging partial convolutions for handling missing or corrupted image regions. This study provides valuable insights into advancing  by utilizing partial convolutions for more visually appealing and accurate results. The researchers' technique shows promising potential for future developments in this field.