一种基于学习的非线性人脸图像超分辨率算法  

Learning-based nonlinear algorithm of face image super-resolution

作　　者：黄东军[1] 侯松林[1] 

机构地区：[1]中南大学信息科学与工程学院,长沙410083

出　　处：《计算机应用》2009年第5期1339-1341,共3页journal of Computer Applications

基　　金：国家自然科学基金资助项目(60873188)

摘　　要：提出了一种单幅人脸图像的超分辨率重构算法。该算法采用马尔可夫网络模型描述重构机制,对输入的低分辨率图像,以及训练用高分辨率图像和对应的低分辨率图像进行分块,并使图像基本对齐,构造训练图像集。针对简化马尔可夫网络计算的需要以及训练集人脸图像的差异,在采用块坐标限位操作的基础上,使用了一种非线性样本搜索算法,降低了搜索空间复杂度,提高了匹配效率和相关性。算法利用搜索到的高分辨率图像分块样本,直接输出超分辨率图像。分析和实验表明,与传统学习算法相比,该方法具有输出质量好、效率高的特点。A super resolution algorithm for single face image based on learned image examples was proposed. The algorithm used patch-based Markov network to express the mechanism of super-resolution processing. After dividing the high- resolution images and the corresponding low-resolution ones into patches, the training dataset was set up. Considering the requirements of Markov network computing and the difference among the images in training dataset, a patch position constraint operation for searching the matched patch and a nonlinear searching algorithm were used. These techniques decreased the complexity of the searching operation and increased the effect of matching. After collecting the matched high-resolution patches, the proposed method directly used them to integrate an output image. Experimental results demonstrate that the approach has better performance and higher efficiency.

关 键 词：人脸图像 超分辨率 马尔可夫网络 非线性搜索 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]