面向图像超分辨率的紧凑型多径卷积神经网络算法研究  

Research on Image Super-resolution Based on Compact Multi-path Convolution Neural Network

作　　者：应自炉[1] 商丽娟 徐颖[1] 刘健 YING Zi-lu;SHANG Li-juan;XU Ying;LIU Jian(School of Information Engineering,Wuyi University,Jiangmen,Guangdong 529020,Chin)

机构地区：[1]五邑大学信息工程学院,广东江门529020

出　　处：《信号处理》2018年第6期668-679,共12页Journal of Signal Processing

基　　金：国家自然科学基金项目(61372193;61771347);广东高校优秀青年教师培训计划资助项目(SYQ2014001);广东省特色创新项目(2015KTSCX143;2015KTSCX145);广东省青年创新项目(2016KQNCX171)

摘　　要：为改善单帧图像分辨率退化问题,减少网络参数,本文提出一种基于紧凑型多径结构卷积神经网络的图像超分辨率重构算法。本文算法采用多径结构模型充分使用低分辨率图像信息,并利用残差学习策略学习低分辨率和高分辨率图像间残差信息以重建高分辨率图像。当卷积核数量有限时,含有Re LU的网络重构性能表现不佳,因此引入最大特征图激活函数,增强网络泛化能力,使网络结构更加紧凑,以捕捉具有竞争性特征,完成图像超分辨率重构。实验结果表明,本文方法具有良好的重构能力,图像清晰度和边缘锐度明显提高,在客观评价和主观视觉效果方面优于当前主流的超分辨率重构方法。为便携式高性能超分辨率重构奠定理论基础。In order to improve the resolution degradation problem of static image and reduce network parameters,a compact multi-path convolution neural network based image super-resolution algorithm is proposed in this paper. This multi-path structure model is adopted to make full use of low-resolution image information,and we use a residual learning strategy to learn residual information between the low level and high level images to reconstruct images super-resolution. Nevertheless,networks with Re LU tend to perform poorly when the number of filter parameters is constrained to a small number. To enhance network generalization ability and make the network more compact,the activated the function named Max-featureMax is adopted here. Experimental results show that the proposed method has good reconstruction ability with image clarity and edge sharpness and has better super-resolution perform in the both objective evaluation and subjective visual effect.This method lays a theoretical foundation for portable high-performance super-resolution reconstruction.

关 键 词：卷积神经网络 超分辨率重构 最大特征图激活函数 多径结构 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]