基于分类预测器及退化模型的图像超分辨率快速重建  

Image fast super-resolution reconstruction based on class predictor and degradation model

作　　者：杨欣[1,2] 费树岷[2] 周大可[1] 唐庭阁[1] 

机构地区：[1]南京航空航天大学自动化学院,南京210016 [2]东南大学自动化学院,南京210096

出　　处：《东南大学学报（自然科学版）》2013年第1期35-38,共4页Journal of Southeast University：Natural Science Edition

基　　金：国家自然科学基金资助项目(60905009;61172135);高等学校博士学科点专项科研基金资助项目(20093218120015);北京师范大学遥感科学国家重点实验室开放基金资助项目(2009KFJJ012);南京航空航天大学基本科研业务费专项科研资助项目(NS2010081)

摘　　要：对基于学习的领域嵌套超分辨率重建方法进行了有效改进,提出了一种基于分类预测器以及退化模型的图像超分辨率重建技术.首先,利用退化模型得到图像训练集,并基于邻域嵌套进行分块;其次,根据图像各自特点提取灰度和梯度特征,并进行特征融合,从而实现了训练过程中噪声信息的有效抑制及图像中边缘信息的锐化;然后,引入分类预测器的思想,设计了一种离线的分类预测器,对预测器进行离线训练,得出优化参数,从而大幅度减少了优化时间;最后,利用L2范数对低分辨率图像分块进行分类,将分块送入相应子预测器中进行快速超分辨率重建.实验结果表明,该算法具有良好的实时性和有效性.Super-resolution （SR） reconstruction technology based on neighbor embedding is effectively improved and a novel image SR reconstruction method using class predictor and degradation model is proposed. First, according to image degradation model, training set is obtained and cut into patches based on neighbor embedding. Secondly, in order to suppress noise and smoothen regions, gray and gradient information is extracted and combined to feature vector according to each patch character. Thirdly, the idea of class predictor is introduced and a novel off-line predictor is designed. Optimal parameters are obtained through off-line training and the optimization time is substantially reduced. Finally, in the light of L2 norm, each low resolution （LR） patch is classed and then put into corresponding sub-predictor with fast SR reconstruction. The experimental results exhibit the good real-time performance and effectiveness of the proposed algorithm.

关 键 词：超分辨率重建 分类预测器 退化模型 特征提取 邻域嵌套 

分 类 号：TH457[机械工程—机械制造及自动化]