基于马尔可夫网络人脸图像超分辨率非线性算法

Markov network-based nonlinear algorithm of face image super-resolution

研究单幅人脸图像的超分辨率重构算法。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,提出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。分析和实验证实,与传统学习算法相比,本方法具有输出质量好、效率高的特点。

This paper researched single face image super resolution algorithms based on learned image examples. It 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, set up the training dataset. Considering the requirements of Markov network computing and the difference among the images in training dataset, it proposed a patch position constraint operation for searching the matched patch and a nonlinear searching algorithm. These techniques could decrease the complexity of the searching operation and increase 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 algorithm has a better performance and higher efficiency.