摘要
深度卷积神经网络已被成功应用于图像高斯白噪声的去除。针对去除真实图像噪声的需要,本文提出了一种基于生成对抗网络的去噪算法。生成网络采用U-net结构,并通过嵌入残差密集块以更好地提取图像特征与减少细节丢失。同时,判别网络采用全卷积网络架构来实现图像的像素级分类,以提升判别器性能。此外,设计了一种增强网络结构,以进一步提高去噪图像质量。仿真实验结果表明,该算法视觉效果以及去噪性能指标PSNR、SSIM均优于其他同类算法,能够更有效地恢复图像细节。
Deep convolution neural network has been successfully applied to remove Gaussian white noise from images.In order to remove the real noise of image,a denoising algorithm based on generative adversarial network is proposed.The generative network adopts U-net structure by embedding residual dense blocks to better extract image features with less loss of details.The adversarial network adopts the full convolution network architecture to realize the pixel level classification of the image so as to improve the performance of the discriminator.An enhanced network structure is designed to further improve the quality of denoised images.The results of subjective and objective evaluations of the experiment show that the visual effects and the denoising performance of PSNR,SSIM of the algorithm are superior to other similar algorithms,and can recover im-age details more effectively.
作者
江巨浪
严华锋
刘国明
JIANG Juang;YAN Huafeng;LIU Guoming(School of Electronic Engineering and Intelligent Manufacturing,Anqing Normal University,Anqing 246133,China)
出处
《安庆师范大学学报（自然科学版）》
2023年第4期74-78,共5页
Journal of Anqing Normal University(Natural Science Edition)
基金
安徽省自然科学基金(2108085MF196)。
关键词
图像去噪
生成对抗网络
真实噪声
残差密集块
image denoising
generative adversarial network
real noise
residual dense block
