摘要
针对现有基于样本学习的人脸超分辨率算法对人脸图像采用全局搜索,存在非局部误匹配且复原图像视觉效果不佳等问题,提出了一种新的基于匹配学习的人脸图像超分辨率算法。首先根据输入图像预分类得到一个样本子类库,并构建相应的特征图像。在匹配过程中,针对不同人脸图像,采用2种新的搜索策略,考虑了图像块之间的相似性和一致性,使复原图像看起来更加连贯自然。实验结果表明,与其他方法相比,本文算法生成的高分辨率人脸图像获得了更好的视觉效果和更高的平均峰值信噪比,具有很好的实用价值。
The exsiting example-based super-resolution algorithms of face image adopt global search, which causes the problems of non-local mismatch and poor visual effect of image restoration. A new matching and learning-based face image super-resolution restoration algorithm is proposed. A pre-classification process of input image is applied to get a sub-sample library from the image library, and the corresponding feature images are created. In the matching process, two new search strategies for different face images are used, which consider the similarity and consistency between image patches and make the recovered image look more coherent and natural. Experimental results show that the proposed algorithm synthesizes high-resolution faces with better visual effect and obtains higher values of the average of Peak Signal-to-Noise Ratios(PSNR) when compared with other methods.
出处
《太赫兹科学与电子信息学报》
2015年第2期291-296,共6页
Journal of Terahertz Science and Electronic Information Technology
基金
国家自然科学基金资助项目(61271330)
关键词
超分辨率
预分类
局部限位搜索
一致搜索
super-resolution
pre-classification
local search
coherence search