中文摘要：

本文针对目前大多数基于学习的超分辨率算法由于＂分类算法＂造成的＂量化＂误差的问题,提出了基于SVR的人脸图像超分辨率算法.算法首先分别提取训练库中高低分辨率图像块的高频信息和中频信息（差分高斯特征,DoG）作为建立回归关系的特征,依据它们的关系（并考虑人脸的特殊性）使用SVR建立起回归模型.在复原时,将待复原的低分辨率图像的中频特征输入已经建立的SVR回归模型得到需要的高频信息.通过对亚洲人脸库（亚洲人为主）IMDB和Yale人脸库（欧美人为主）的实验结果表明,本文提出的方法对亚洲人脸和欧美人脸都能取得了较好的复原效果,复原的图像在主观的视觉效果和客观的峰值信噪比上都取得较好的结果.

英文摘要：

Most learning-based super-resolution algorithms have the shortcoming of the＂quantitative＂ er- ror when using classification algorithm. In this paper, a learning based super-resolution algorithm based on SVR is proposed. The algorithm first extracts high-frequency information of High-resolution images and middle-frequency of low-resolution images. Then, according to their relationship, a regression mod- el is built by using SVR. During the recovery, middle-frequency of low-resolution images is extracted to feed into the built model to get high-frequency information. The experimental results showed that our method achieves very good results to IMDg（Asians） face database and Yale face（mainly Europeans and Americans） database. Overall, the results of our method have better visual effects and higher peak sig- nal to noise ratio.