摘要
针对目前大多数基于深度学习的图像超分辨率重建方法都不考虑尺度与几何变化的问题,提出基于双层可变形卷积网络的超分辨率图像重建方法。首先,该方法将标准的卷积层替换为可变形卷积层,模拟图像中的简单几何变化过程;其次,利用两个不同尺寸的可变形卷积层构造双层可变形卷积单元,来提取图像在不同尺度下的特征信息;最后,在特征图之间增加残差连接,缓解梯度消失带来的训练难度。实验结果说明该方法比现有的一些重建方法能更好地提取图像的特征信息,提高图像的重建效果。
For the fact that most deep learning based super-resolution reconstruction methods do not consider the variations of scale and geometry,a bilayer deformable convolutional network based super-resolution reconstruction was proposed.Firstly,the standard convolutional layer was replaced with a deformable convolutional layer to simulate the process of simple geometric deformation.Secondly,a bilayer deformable convolution unit was constructed using two different sizes of deformable convolution kernels to extract the features with different scales.Finally,residual connections were added among feature maps to alleviate the network training problem caused by the disappearance of gradient.The experimental results show that our proposed method can extract feature information of the image better to improve the reconstructed effect than some other existing reconstruction methods.
作者
黄陶冶
赵建伟
周正华
HUANG Taoye;ZHAO Jianwei;ZHOU Zhenghua(Department of Information and Mathematics,China Jiliang University,Hangzhou Zhejiang 310018,China)
出处
《计算机应用》
CSCD
北大核心
2019年第S02期68-74,共7页
journal of Computer Applications
基金
国家自然科学基金资助项目(61571410)
浙江省自然科学基金资助项目(LY18F020018,LSY19F020001)
关键词
超分辨率重建
双层可变形卷积
残差学习
几何变化
多尺度
super-resolution reconstruction
bilayer deformable convolution
residual learning
geometric deformation
multi-scale
