摘要
为进一步改善生成对抗网络重建图像的视觉效果,针对网络模型训练不稳定、收敛困难等因素,提出一种改进的生成对抗网络的图像超分辨率算法。用Wasserstein代替JS散度优化网络,稳定网络的训练,加入优化残差块,深化网络高频特征提取,去除残差块中的冗余层保证精度的同时减少计算。实验结果表明,该方法的网络结构训练稳定,图像的评价指标相较于其它几种对比方法有所提升,重建图像纹理丰富,视觉效果逼真。
To further improve the visual effects of reconstructed images generated through the generative adversarial network,an improved image super-resolution algorithm for generating the generative adversarial network was proposed in view of the unstable training of the network models and the difficulty of convergence.Wasserstein was used instead of JS divergence to optimize the training of network stability network.Optimized residual blocks were added to deepen the high frequency feature extraction of network.Redundant layers in residual blocks were removed to ensure accuracy and reduce calculation.Experimental results show that the network structure training of the method is stable,and the evaluation index of the image is improved compared with other contrast methods.The reconstructed image has rich texture and realistic visual effects.
作者
王冬冬
王力
姜敏
王可新
栾浩
WANG Dong-dong;WANG Li;JIANG Min;WANG Ke-xin;LUAN Hao(College of Big Data and Information Engineering,Guizhou University,Guiyang 550025,China;School of Information Engineering,Guizhou University of Engineering Science,Bijie 551700,China)
出处
《计算机工程与设计》
北大核心
2020年第7期1981-1986,共6页
Computer Engineering and Design
基金
贵州省教育厅创新群体重大研究基金项目(黔教合KY字[2016]057)
国家新工科实践基金项目(黔教高涵[2018]209号)。
关键词
超分辨率
图像重建
视觉效果
深度学习
生成对抗网络
super-resolution
image reconstruction
visual effect
deep learning
generative adversarial network
