基于二维稀疏表示的人脸超分辨率重构算法

Face Super-Resolution Reconstruction Algorithm Based on Two-Dimensional Sparse Representation

在分析人脸超分辨率算法和二维稀疏表示的基础上,提出基于二维稀疏表示的人脸超分辨率重构算法。与一维稀疏表示中将图像块转换为列向量不同,本文考虑到二维图像列与列之间的近邻关系,对图像块进行二维稀疏表示;在字典训练中,对每组图像块的每一列训练高、低分辨率字典,提出二维K-SVD算法对字典进行训练,减少字典训练消耗的时间,同时能够改善超分辨率人脸的质量。采用中科院CAS-PEAL共享人脸图像数据库进行仿真实验,实验结果从主、客观质量均验证了本文算法的有效性及先进性。

In this paper, by analyzing the face super-resolution algorithms and two-dimensional sparse representation, a novel algorithm called the face super-resolution algorithms based on two- dimensional sparse representation is proposed. Unlike the traditional sparse representation converting image blocks to column vector, the algorithm takes two-dimensional sparse representation with blocks under the constraints that columns in each block own the neighbor-relations. A novel K-SVD algorithm called two-dimensional K-SVD algorithm is proposed to train sparse dictionaries by training high and low resolution dictionaries for each columns of blocks in each set. The two- dimensional K-SVD algorithm can not only reduce the time of the dictionary training effectively, but also improve the quality of the reconstruction of super-resolution images. Experiment results on CAS-PEAL face database show that the algorithm is effective on subjective and objective qualities.