基于双正则化参数的在线字典学习超分辨率重建  

Image super-resolution based on online dictionary learning with two regularization parameters

作　　者：倪浩[1] 阮若林[2] 刘芳华[1] 

机构地区：[1]湖北科技学院电子与信息工程学院,湖北咸宁437100 [2]湖北科技学院生物医学工程学院,湖北咸宁437100

出　　处：《计算机应用研究》2016年第3期911-915,共5页Application Research of Computers

基　　金：国家自然科学基金资助项目(61271256);湖北省高等学校优秀中青年科技创新团队计划项目(T201513);湖北省自然科学基金资助项目(2015CFB452)

摘　　要：基于学习的单图超分辨率重建算法能获得较好的超分效果,但存在重建图像伪影较为明显的问题。为解决这一问题,提出了一种基于双正则化参数的在线字典学习超分辨率重建算法。在字典学习过程中运用在线字典学习方法(online dictionary learning,ODL),并在稀疏字典生成阶段和图像重建阶段分别设置了两个不同的正则化参数。实验中生成的目标高分辨率图像PSNR比经典的稀疏编码超分方法(sparse coding super-resolution,SCSR)平均提高了0.39 d B,在较好地恢复图像边缘锐度和纹理细节的同时有效地抑制了伪影。ODL和双正则化参数的引入,提高了字典训练的精度,使字典训练和图像重建阶段的稀疏系数独立可调,实验中能够有效地消除伪影,提升了超分辨率重建的效果。The performance of some learning-based super-resolution methods are promising,but some obvious artifacts appear in the reconstruction images. In order to solve this problem,this paper presented a novel super-resolution algorithm based on online dictionary learning（ ODL） with two regularization parameters. It employed ODL in the dictionary learning procedure.Then the algorithm set two regularization parameters in the procedures of dictionary learning and image reconstruction. In the experiments,the PSNRs of the new method were 0. 39 d B higher than the state-of-the-art sparse coding super-resolution（ SCSR） in average. It could eliminate the artifacts while recovering the edge sharpness and the texture details efficiently. With the introduction of ODL and two regularization parameters,it promoted the dictionary training accuracy and made the sparse coefficients in dictionary learning and image reconstruction adjustable separately. The experiments show that the artifacts are eliminated effectively. It promotes the final effect of super-resolution reconstruction well.

关 键 词：正则化参数 超分辨率 在线字典学习 稀疏编码 图像 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]