基于稀疏表示和多成分字典学习的超分辨率重建  

Super-resolution reconstruction based on sparse representation and multicomponent dictionaries learning

作　　者：徐志刚[1] 李文文[1] 袁飞祥 朱红蕾[1] 许亚美 XU Zhigang;LI Wenwen;YUAN Feixiang;ZHU Honglei;XU Yarnei(School of Computer and Communication, Lanzhou University of Technology, Lanzhou 730050, Chin)

机构地区：[1]兰州理工大学计算机与通信学院,甘肃兰州730050

出　　处：《系统工程与电子技术》2018年第3期699-703,共5页Systems Engineering and Electronics

基　　金：国家自然科学基金(61761028;61562058);兰州理工大学博士科研基金资助课题

摘　　要：针对目前基于稀疏表示模型的图像超分辨率重建方法对于边缘、纹理等细节信息保持能力有限、易产生视觉伪影的问题,提出了基于稀疏表示和多成分字典学习的超分辨率重建算法。在字典训练阶段,所提算法在利用图像形态分量分析方法构造纹理和结构字典的基础上,为了有效地提取低分辨率图像特征细节信息,对图像结构分量采用一阶二阶导数进行特征提取,对纹理分量采用Gabor变换进行特征提取,并使用L1/2范数构造训练字典模型;而在重建阶段,为了消除重建图像块效应及模糊伪影,进一步提高重建图像的质量,采用全局约束和非局部相似性约束相结合的方法对重建高分辨率图像进行优化。实验结果表明,该算法在重建图像主观和客观评价指标方面均有较好的表现。In order to overcome the disadvantage of some super-resolution methods based on sparse representation that it is not easy to preserve edge and texture detail information and easy to produce visual artifacts,an algorithm based on sparse representation and multicomponent dictionaries learning is proposed.In the stage of dictionary training,the method of morphological component analysis is used to train texture and piecewise smooth dictionaries.In order to extract the detail information of low-resolution images more effectively,the first-order and second-order derivatives are used to extract the piecewise smooth parts.And the texture parts are extracted by the Gabor transform.The L1/2 regularization is used to train the learning dictionaries.In the reconstruction stage,in order to remove blur and minimize visual blocking artifacts,the global constraint term and the nonlocal similarity constraint term are used to optimize the reconstructed high-resolution image.Experimental results can achieve more competitive reconstruction quality of the proposed algorithm both visually and quantitatively.

关 键 词：超分辨率 稀疏表示 L1/2正则化 非局部相似性 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]