Mask Inpaint Using Downsampled Fast Fourier Convolution

EasyChair Preprint 13732

Abstract

In recent years, image inpainting technology has made significant progress. Notably, the LaMa model, proposed in 2021, has shown marked improvements for large-area inpaints. However, challenges persist in effectively inpainting complex geometric structures, processing high-resolution images quickly, and achieving realism in filled areas. Existing solutions to enhance realism often involve incorporating diffusion models to generate the filled mask portions, which increases hardware demands and is impractical for high-resolution images. We propose a novel image inpainting approach called Downsampled Fast Fourier Convolution (DFFC), with the main components being: i) deep learning-based image downsampling, ii) an image inpainting architecture based on Fast Fourier Convolution (FFC), iii) a high-perceptual-domain loss function, and iv) dynamic large-area mask training. Our technique maintains the original model's performance while enhancing processing speed and reducing computational load.

Keyphrases: Fast Fourier Convolution, Inpaint, computer vision

@booklet{EasyChair:13732,
  author    = {Xiaoyang Gao and Tao Yang},
  title     = {Mask Inpaint Using Downsampled Fast Fourier Convolution},
  howpublished = {EasyChair Preprint 13732},
  year      = {EasyChair, 2024}}