摘要
为了更好地去除图像中的噪声,提出了一种改进的深度卷积神经网络(Dncnn)图像去噪算法。针对现有的Dncnn网络参数量大,对Dncnn网络的第2~16层进行了改进,使网络参数量降低1/3后,仍能保持和Dncnn一样的训练效果。在此基础上,对网络底层的低级语义信息和高层的高级语义信息进行了特征融合,使得网络训练更平稳,并能达到更好的训练效果。实验结果表明无论与图像去噪领域公认最好的去噪算法BM3D相比,还是与深度学习领域先进的图像去噪算法Dncnn相比,改进的Dncnn都有更好的表现。
In order to better remove the noise in the image,an improved Dncnn image denoising algorithm was proposed.In view of the large amount of existing Dncnn network parameters,the 2nd^16th layers of the Dncnn network had been improved,and the network parameters were reduced by one-third,and the same training effect as Dncnn can be maintained.On this basis,the low-level semantic information of the underlying network and the high-level semantic information of the upper layer are combined to make the network training more stable and achieve better training results.The experimental results show that the improved Dncnn has better performance than the best denoising algorithm BM3D in the field of image denoising and the advanced image denoising algorithm Dncnn in the depth learning field.
作者
白瑞君
李众
张启尧
刘方涛
BAI Rui-jun;LI Zhong;ZHANG Qi-yao;LIU Fang-tao(School of Software,North University of China,Taiyuan 030051,China)
出处
《科学技术与工程》
北大核心
2019年第36期247-252,共6页
Science Technology and Engineering
关键词
图像去噪
网络参数
低级语义信息
高级语义信息
特征融合
image denoising
network parameters
low-level semantic information
high-level semantic information
feature fusion
