摘要
文章提出了一种基于卷积神经网络结构的全景图像超分辨率算法3D-WDSR,在单帧图像超分网络WDSR基础上引入可变卷积核结构以减小参数量,节省计算资源。实验结果表明,在不同尺度的超分辨率任务中,先经过预训练后的网络具有更好的性能表现,所提出的3D-WDSR算法的超分辨率重建效果要高于双三次插值方法和EDSR算法,且在参数量仅为WDSR网络的22.3%的情况下具有相当的超分辨率性能。
This paper proposes a panoramic image super-resolution algorithm 3D-WDSR which is based on the singleframe image super-division network WDSR,adding a variable convolution kernel structure to reduce the amount of parameters and to save computing resources.Experimental results show that in super-resolution tasks of different scales,the pre-trained network has better performance.At the same time,the super-resolution reconstruction effect of the 3D-WDSR network is higher than the bilinear interpolation method and the EDSR algorithm,and has considerable super-resolution performance when the parameter amount is only 22.3%of the WDSR network.
作者
董桂官
吴双彤
张汉琦
DONG Gui-guan;WU Shuang-tong;ZHANG Han-qi(China Electronics Standardization Institute,Beijing 100176,China;School of Information and Electronic,Beijing Institute of Technology,Beijing 100081,China)
出处
《电脑与信息技术》
2022年第2期1-4,共4页
Computer and Information Technology
基金
国家自然科学基金项目(项目编号:61620106002)。
关键词
深度学习
卷积神经网络
全景图像
可变卷积核
超分辨率
Deep learning
Convolutional neural network
Panoramic image
Variable convolution kernel
Super-resolution