Paper Image

Dynamic learning of attention and convolution for image restoration

Published on:

9 November 2023

Primary Category:

Computer Vision and Pattern Recognition

Paper Authors:

Kui Jiang,

Xuemei Jia,

Wenxin Huang,

Wenbin Wang,

Zheng Wang,

Junjun Jiang

Bullets

Key Details

Proposes association learning to combine advantages of self-attention and convolution for image restoration

Designs a multi-input attention module to relate rain removal and background recovery

Constructs a hybrid network fusing Transformer and CNN branches

Achieves state-of-the-art results on image deraining, underwater and low-light enhancement

AI generated summary

Dynamic learning of attention and convolution for image restoration

This paper proposes an association learning method that combines self-attention and convolution to utilize their advantages and suppress disadvantages, achieving high-quality and efficient image restoration. The key is a multi-input attention module that associates rain streak removal and background recovery by generating a degradation prior. This allows full use of background textures from the rainy image to aid recovery. A hybrid network with Transformer and CNN branches is designed to represent global structure and local details. Experiments on image deraining, underwater enhancement, and low-light enhancement demonstrate superiority over state-of-the-art methods.

Answers from this paper

Comments

No comments yet, be the first to start the conversation...

Sign up to comment on this paper

Sign Up