通信学报 ›› 2021, Vol. 42 ›› Issue (3): 229-237.doi: 10.11959/j.issn.1000-436x.2021049

• 学术通信 • 上一篇    

基于多通道GAN的图像去噪算法

王洪雁1,2,3, 杨晓2, 姜艳超2, 汪祖民2   

  1. 1 浙江理工大学信息学院,浙江 杭州 310018
    2 大连大学信息工程学院,辽宁 大连 116622
    3 五邑大学智能制造学部,广东 江门 529020
  • 修回日期:2021-01-28 出版日期:2021-03-25 发布日期:2021-03-01
  • 作者简介:王洪雁(1979– ),男,河南南阳人,博士,浙江理工大学特聘教授、硕士生导师,主要研究方向为稀疏学习、阵列信号处理、参数估计、机器视觉等。
    杨晓(1997– ),女,山东德州人,大连大学硕士生,主要研究方向为图像处理、机器视觉等。
    姜艳超(1985– ),女,河北唐山人,博士,大连大学硕士生导师,主要研究方向为信号处理、机器学习等。
    汪祖民(1975– ),男,河南信阳人,博士,大连大学教授、硕士生导师,主要研究方向为信号处理、机器学习等。
  • 基金资助:
    国家自然科学基金资助项目(61301258);国家自然科学基金资助项目(61871164);浙江省自然科学基金重点资助项目(LZ21F010002);中国博士后科学基金资助项目(2016M590218)

Image denoising algorithm based on multi-channel GAN

Hongyan WANG1,2,3, Xiao YANG2, Yanchao JIANG2, Zumin WANG2   

  1. 1 School of Information Science and Technology, Zhejiang Sci-Tech University, Hangzhou 310018, China
    2 College of Information Engineering, Dalian University, Dalian 116622, China
    3 Faculty of Intelligent Manufacturing, Wuyi University, Jiangmen 529020, China
  • Revised:2021-01-28 Online:2021-03-25 Published:2021-03-01
  • Supported by:
    The National Natural Science Foundation of China(61301258);The National Natural Science Foundation of China(61871164);Key Projects of Natural Science Foundation of Zhejiang Province(LZ21F010002);China Postdoctoral Science Foundation(2016M590218)

摘要:

针对图像采集和传输过程中所产生噪声导致后续图像处理能力下降的问题,提出基于生成对抗网络(GAN)的多通道图像去噪算法。所提算法将含噪彩色图像分离为RGB三通道,各通道基于具有相同架构的端到端可训练的GAN实现去噪。GAN生成网络基于U-net衍生网络以及残差块构建,从而可参考低级特征信息以有效提取深度特征进而避免丢失细节信息;判别网络则基于全卷积网络构造,因而可获得像素级分类从而提升判别精确性。此外,为改善去噪能力且尽可能保留图像细节信息,所构建去噪网络基于对抗损失、视觉感知损失和均方误差损失这3类损失度量构建复合损失函数。最后,利用算术平均方法融合三通道输出信息以获得最终去噪图像。实验结果表明,与主流算法相比,所提算法可有效去除图像噪声,且可较好地恢复原始图像细节。

关键词: 图像去噪, 生成对抗网络, 通道分离, 复合感知损失

Abstract:

Aiming at the issue that the noise generated during image acquisition and transmission would degrade the ability of subsequent image processing, a generative adversarial network (GAN) based multi-channel image denoising algorithm was developed.The noisy color image could be separated into red-green-blue (RGB) three channels via the proposed approach, and then the denoising could be implemented in each channel on the basis of an end-to-end trainable GAN with the same architecture.The generator module of GAN was constructed based on the U-net derivative network and residual blocks such that the high-level feature information could be extracted effectively via referring to the low-level feature information to avoid the loss of the detail information.In the meantime, the discriminator module could be demonstrated on the basis of fully convolutional neural network such that the pixel-level classification could be achieved to improve the discrimination accuracy.Besides, in order to improve the denoising ability and retain the image detail as much as possible, the composite loss function could be depicted by the illustrated denoising network based on the following three loss measures, adversarial loss, visual perception loss, and mean square error (MSE).Finally, the resultant three-channel output information could be fused by exploiting the arithmetic mean method to obtain the final denoised image.Compared with the state-of-the-art algorithms, experimental results show that the proposed algorithm can remove the image noise effectively and restore the original image details considerably.

Key words: image denoising, generative adversarial network, channel separation, joint perception loss

中图分类号: 

No Suggested Reading articles found!