基于多类预测器学习的超分辨率复原算法  

Super-resolution restoration algorithm based on multi-class prediction learning

作　　者：汤嘉立 柳益君[1] 杜卓明 TANG Jia-li;LIU Yi-jun;DU Zhuo-ming(College of Computer Engineering,Jiangsu University of Technology,Changzhou 213001,China;School of Mechanical Engineering,Jiangsu University,Zhenjiang 212013,China;School of Mathematical Science,Nanjing Normal University,Nanjing 210023,China)

机构地区：[1]江苏理工学院计算机工程学院,江苏常州213001 [2]江苏大学机械工程学院,江苏镇江212013 [3]南京师范大学数学科学学院,江苏南京210023

出　　处：《计算机工程与设计》2018年第4期1100-1105,共6页Computer Engineering and Design

基　　金：国家自然科学基金项目(61402206);中国博士后科学基金项目(2016M601845);住房城乡建设部研究开发基金项目(2016-K8-028)

摘　　要：为解决样本学习超分辨率算法的图像样本误匹配和边缘平滑问题,提出一种基于神经网络的非线性多类预测器学习算法,设计神经网络多类预测器,采用小生境基因表达式编程方法优化反向传播神经网络。通过学习样本集对预测器进行训练,学得学习样本中的先验知识,根据从低分辨率图像块提取的特征矢量预测图像高频信息,完成图像超分辨率复原。实验结果表明,相比样本预分类学习的几种算法,该算法的PSNR和SSIM值均有了一定提升,主观上复原结果具有更丰富的细节。To solve the problems of image mismatching and smooth edges in pre-classified based super-resolution algorithms,a non-linear multi-class prediction learning algorithm based on neural network was proposed.The neural network multi-class predictor was designed,and niche gene expression programming was used to optimize the back-propagation neural network.The predictor was trained by learning samples and prior knowledge of samples was obtained to predict high frequency information of image and further to complete the super-resolution image restoration.Experimental results show that compared with sample pre-classified based algorithms,PSNR and SSIM of the proposed algorithm are improved respectively,and subjectively the restoration results are more abundant in detail.

关 键 词：超分辨率复原 多类预测器 小生境算法 基因表达式编程 神经网络 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]