摘要
SRGAN算法虽然具有许多优点,但也存在图像重建效果不够好、参数数量庞大、激活函数表现较差等问题。为此,本文提出一种基于SRGAN的图像超分辨率算法SRGAN-E。该算法首先删除BN层,提高图形的重建效果;再在原生成器模型中加入一维卷积注意力机制,使得图像在重建过程中更加关注上下文信息并减少网络模型中生成器的参数;将SRGAN算法鉴别器模型中的LeakyReLU函数改为Mish函数,以提升鉴别器的性能。实验结果表明:对比SRGAN算法,改进后的SRGAN-E算法在4个测试集上PSNR的平均值增加了0.345,SSIM的平均值增加了0.009;SRGAN-E算法的生成器参数数量与SRGAN算法相比,减少了1 388个。SRGAN-E算法不但提高了图像的重建效果而且还减少了模型参数。
Although SRGAN algorithm has many advantages,it also has some problems such as low reconstruction effect,large number of parameters and poor performance of activation function.Therefore,an image super-resolution algorithm SRGAN-E based on SRGAN is proposed in this paper.Firstly,BN layer is deleted to improve the reconstruction effect.Furthermore,a one-dimensional convolutional attention mechanism was added to the original generator model to make the image pay more attention to the context information and reduce the generator parameters in the network model.To improve the performance of the discriminator,the LeakyReLU function in the discriminator model of SRGAN algorithm was changed to Mish function.The experimental results show that compared with SRGAN algorithm,the average value of PSNR and SSIM of SRGAN-E algorithm are increased by 0.345 and 0.009 respectively on four test sets.The number of generator parameters of SRGAN-E algorithm is reduced by 1 388 compared with SRGAN algorithm.The SRGAN-E algorithm not only improves the image reconstruction effect but also reduces the model parameters.
作者
刘嵩山
王华军
李特
李光志
万军
马瑜
LIU Songshan;WANG Huajun;LI Te;LI Guangzhi;WAN Jun;MA Yu(School of Computer and Network Security,Chengdu University of Technology,Chengdu 610059 China)
出处
《西华大学学报（自然科学版）》
CAS
2022年第5期84-90,103,共8页
Journal of Xihua University:Natural Science Edition
基金
四川省人工智能重点实验室项目(2020RYJ02)
模式识别与智能信息处理四川省高校重点实验室项目(MSSB-2020-10)。
关键词
卷积神经网络
深度学习
超分辨率算法
激活函数
图像重建
convolutional neural network
deep learning
super-resolution algorithm
activation function
image reconstruction
作者简介
通信作者:王华军(1964-),教授,博士,博士生导师,主要研究方向为图像超分辨率。ORCID:0000-0001-5897-5562,E-mail:wanghuajun@cdut.edu.cn。
