文献类型：期刊文献

英文题名：TMS-GAN: A Twofold Multi-Scale Generative Adversarial Network for Single Image Dehazing

作者：Wang, Pengyu[1];Zhu, Hongqing[1];Huang, Hui[2];Zhang, Han[1];Wang, Nan[1]

机构：[1]East China Univ Sci & Technol, Sch Informat Sci & Engn, Shanghai 200237, Peoples R China;[2]Shanghai Normal Univ, Coll Informat & Elect Engn, Shanghai 200237, Peoples R China

年份：2022

卷号：32

期号：5

起止页码：2760

外文期刊名：IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS FOR VIDEO TECHNOLOGY

收录：;EI(收录号：20213010675118);WOS:【SCI-EXPANDED(收录号:WOS:000790830300024)】；

基金：This work was supported by the National Nature Science Foundation of China under Grant 61872143.

语种：英文

外文关键词：Atmospheric modeling; Training; Scattering; Image restoration; Generative adversarial networks; Decoding; Image color analysis; Single image dehazing; twofold; multi-scale; fake realistic haze generation; multi-attention progressive fusion

摘要：In recent years, learning-based single image dehazing networks have been comprehensively developed. However, performance improvement is limited due to domain shift between trained synthetic hazy images and untrained real-world hazy images. To alleviate this issue, this paper proposes a real-world dehazing targeted training scheme which nearly realizes paired real-world data training. As a result, a Twofold Multi-scale Generative Adversarial Network (TMS-GAN) consisting of a Haze-generation GAN (HgGAN) and a Haze-removal GAN (HrGAN) is designed. HgGAN attributes real haze properties to synthetic images and HrGAN removes haze from both synthetic and generated fake realistic data under supervision. Thus, the proposed method can better adapt to real-world image dehazing using this cooperative training scheme. Meanwhile, several structural advances of TMS-GAN also improve dehazing performance. Specifically, a haze residual map based on atmospheric scattering model is deduced in HgGAN for fake realistic data generation. The dual-branch generator in HrGAN draws attention to detail restoration by one branch along with another color-branch. A plug-and-play Multi-attention Progressive Fusion Module (MAPFM) is proposed and inserted in both HgGAN and HrGAN. MAPFM incorporates multi-attention mechanism to guide multi-scale feature fusion in a progressive manner, in which Adjacency-attention Block (AAB) can capture contributing features of each level and Self-attention Block (SAB) can establish non-local dependency of feature fusion. Experiments on mainstream benchmarks show that the proposed framework is superior especially on real-world hazy images among single image dehazing methods.