通信学报 ›› 2022, Vol. 43 ›› Issue (10): 210-222.doi: 10.11959/j.issn.1000-436x.2022201

• 学术通信 • 上一篇    下一篇

四通道无监督学习图像去雾网络

刘威1,2, 陈成1,2, 江锐1,2, 卢涛1,2   

  1. 1 武汉工程大学计算机科学与工程学院,湖北 武汉 430205
    2 武汉工程大学智能机器人湖北省重点实验室,湖北 武汉 430205
  • 修回日期:2022-09-08 出版日期:2022-10-25 发布日期:2022-10-01
  • 作者简介:刘威(1987− ),男,湖北荆门人,博士,武汉工程大学讲师、硕士生导师,主要研究方向为计算机视觉、图像处理、深度学习等
    陈成(1998− ),男,贵州晴隆人,武汉工程大学硕士生,主要研究方向为图像处理、深度学习
    江锐(1998− ),男,湖北黄冈人,武汉工程大学硕士生,主要研究方向为图像处理、深度学习
    卢涛(1980− ),男,湖北武汉人,博士,武汉工程大学教授、博士生导师,主要研究方向为计算机视觉、图像处理、模式识别等
  • 基金资助:
    国家自然科学基金资助项目(62001334);国家自然科学基金资助项目(62072350);湖北省支持企业技术创新发展基金资助项目(2021BLB172)

Four-path unsupervised learning-based image defogging network

Wei LIU1,2, Cheng CHEN1,2, Rui JIANG1,2, Tao LU1,2   

  1. 1 School of Computer Science and Engineering, Wuhan Institute of Technology, Wuhan 430205, China
    2 Hubei Key Laboratory of Intelligent Robot, Wuhan Institute of Technology, Wuhan 430205, China
  • Revised:2022-09-08 Online:2022-10-25 Published:2022-10-01
  • Supported by:
    The National Natural Science Foundation of China(62001334);The National Natural Science Foundation of China(62072350);Technology Innovation and Development Project of Supports Enterprise of Hubei Province(2021BLB172)

摘要:

摘 要:为了解决单幅图像去雾领域有监督网络和无监督网络分别存在的问题,基于循环生成对抗网络(CycleGAN),提出了一种四通道无监督学习图像去雾网络。所提模型主要包括3个子网络,即去雾网络、合成雾网络和注意力特征融合网络;并由此3个子网络顺序组合构建了4条学习通道,即去雾通道、去雾结果颜色-纹理恢复通道、合成雾通道以及合成雾结果颜色-纹理恢复通道。特别地,在合成雾网络中,为了更好地约束去雾网络生成更高质量的无雾图像,引入了大气散射模型(ASM)以加强网络从有雾图像域到无雾图像域的映射转换;同时,为了进一步提高去雾网络和合成雾网络的图像生成质量,提出了一种注意力特征融合网络,该网络基于雾相关派生图,采用多路映射结构和注意力机制,加强对生成图像颜色、纹理细节等信息的恢复。在合成雾和真实雾图数据集上的大量实验结果表明,所提方法能更好地恢复各类场景中雾图的颜色和纹理细节等信息。

关键词: 循环生成对抗网络, 单幅图像去雾, 大气散射模型, 注意力特征融合, 无监督学习

Abstract:

To solve the problems of supervised network and unsupervised network in the field of single image defogging, a four-path unsupervised learning-based image defogging network based on cycle generative adversarial network (CycleGAN) was proposed, which mainly included three sub-networks: defogging network, synthetic fog network and attention feature fusion network.The three sub-networks were sequentially combined to construct four learning paths, which were the defogging path, the color-texture recovery path for defogged result, the synthetic fog path, and the color-texture recovery path for synthetic fog result.Specifically, in the synthetic fog network, to better constrain the defogging network to generate higher quality fogfree images, the atmospheric scattering model (ASM)was introduced to enhance the mapping transformation of the network from the foggy image domain to the fogfree image domain.Furthermore, to further improve the image generation quality of the defogging network and the synthetic fog network, an attention feature fusion network was proposed.The proposed network was based on several fog-derived images, which adopts a multi-channel mapping structure and an attention mechanism to enhance the recovery of color and texture details.Extensive experiments on both synthetic and real-world datasets show that the proposed method can better restore the color and texture details information of foggy images in various scenes.

Key words: cycle generative adversarial network, single image defogging, atmospheric scattering model, attention feature fusion, unsupervised learning

中图分类号: 

No Suggested Reading articles found!