对LLE图像超分辨率算法的几个改进

Several Improvements to LLE Image Super Resolution Algorithm

本文旨在对一种经典的图像超分辨率方法--LLE算法(局部线性嵌入算法)及其代码进行一些改进和优化。为提高对大量图像块的搜索速度,我们采用kd树算法整理样本集;鉴于像素点灰度值的非负性,我们采用非负最小二乘法而不是LLE原来的最小二乘法,确定低分辨率图像块与训练样本集中k最邻近图像块的回归系数;最后,考虑到样本集选取和变换会对实验结果造成影响,我们对训练样本图像的若干因素进行一系列组合,通过正交实验设计得出样本集的最佳选取标准。实验表明,改进后的LLE图像超分辨率算法相比传统的图像插值算法和原LLE算法,效果有较大的改进。

This paper concerns of the improvement and optimization of the code of a classical example based single image super-resolution method:LLE algorithm(locally linear embedding algorithm).For improving the speed of searching large amount of image patches,we employ the kd-tree algorithm to the example patches of training set;furthermore instead of employing the least square regression method,we apply non-negative version of it to obtain the regression coefficients,based on the factor that the non-negative of pixel gray values of image;at last,considering the important effect of the selection and translation of examples on the results of super-resolution methods,we employ the orthogonal test design to some factors of image affine translation for example selection of training set.Experiments demonstrated that our new LLE algorithm can improve the performance greatly,comparing to conventional image interpolation methods and the original LLE method.