电子学报 ›› 2021, Vol. 49 ›› Issue (11): 2160-2165.DOI: 10.12263/DZXB.20200822

• 学术论文 • 上一篇    下一篇

一种基于ARD‑GAN的低照度图像增强方法

江泽涛1, 钱艺1, 伍旭1, 张少钦2   

  1. 1.桂林电子科技大学广西图像图形与智能处理重点实验室,广西 桂林 541004
    2.南昌航空大学,江西 南昌 330063
  • 收稿日期:2020-08-03 修回日期:2021-01-15 出版日期:2021-11-25 发布日期:2021-11-25
  • 作者简介:江泽涛 男, 1961年3月出生,江西九江人.桂林电子科技大学教授,博士生导师,主要从事图像处理、计算机视觉、信息安全方面的研究.
    钱 艺(通讯作者) 男, 1993年6月出生,山东淄博人.桂林电子科技大学硕士研究生,主要从事图像增强方面的研究.E‑mail:Qyizos@163.com
  • 基金资助:
    国家自然科学基金(61876049);广西科技计划(AC16380108);广西图像图形智能处理重点实验(GIIP2006);广西研究生教育创新计划(2019YCXS043);桂林电子科技大学研究生科研创新(2020YCXS050)

Low‑Light Image Enhancement Method Based on ARD‑GAN

Ze-tao JIANG1, Yi QIAN1, Xu WU1, Shao‑qin ZHANG2   

  1. 1.The Key Laboratory of Image and Graphic Intelligent Processing in Guangxi,Guilin University of Electronic Technology,Guilin,Guangxi 541004,China
    2.Nanchang Hangkong University,Nanchang,Jiangxi 330063,China
  • Received:2020-08-03 Revised:2021-01-15 Online:2021-11-25 Published:2021-11-25

摘要:

为解决低照度图像增强过程中噪声放大、细节不足以及色彩还原问题,本文提出一种基于注意力机制残差密集生成对抗网络(Attention Residual Dense?Generative Adversarial Networks, ARD?GAN)的低照度图像增强方法.首先,该方法在全局光照估计模块(Global Illumination Estimation Module, GIEM)中生成全局曝光注意力图,以引导后续模块更好地进行照度增强;其次,使用卷积残差模块(Convolution and Residual Module, CRM)和基于通道注意力的残差密集模块(Channel Attention Residual Dense Module, CARDM)分别提取浅层特征和深层特征,并将不同层次的特征融合以获取更好的细节信息;然后,在CARDM基础上将密集连接与批归一化相结合抑制噪声;最后改进了损失函数,使增强后图像色彩还原更好.实验表明,ARD?GAN有与主流算法相比,在主观视觉和客观评价指标上均得到更好的效果.

关键词: 低照度增强, 图像细节增强, 降噪, 色彩还原, 注意力机制, 残差密集网络

Abstract:

In order to solve the problems of noise amplification, insufficient detail and color restoration in the process of low?light image enhancement, this paper proposes a method of low?light image enhancement based on attention residual dense?generative adversarial networks (ARD?GAN). Firstly, the method generates a global exposure attention map in the global illumination estimation module (GIEM) to guide the subsequent modules to enhance the illumination better; secondly, it adopts the convolution and residual module (CRM) and the channel attention residual dense module (CARDM) to extract shallow features and deep features respectively, and fuses different features of levels to obtain better detailed information; furthermore, based on the CARDM, the dense connection and batch normalization are combined to suppress noise. Finally, the improved loss function restores the enhanced image color better. Comprehensive experiments are conducted, which show that ARD?GAN can significantly outperform mainstream algorithms in subjective vision and objective evaluation indicators.

Key words: low?light enhancement, image detail enhancement, noise reduction, color reproduction, attention mechanism, residual dense network

中图分类号: