摘要
提出一个单幅人脸图像的超分辨率重构算法。该算法建立在马尔可夫网络模型的基础上,引入了语义相似度的学习,将学习的范围限定在位置相关的特征语义区域,提升了学习算法的效率以及重构图像时的逼真性;重构算法中引入了权值融合机制,提升了输出图像的高频成分,有效地改善了图像的全局效果。分析和实验表明,该算法能在大容量训练集中,快速学习到有价值的图像信息,并且在图像的复原的过程中有效地抑制了图像失真现象,极大地改善了超分辨率图像的质量。
A super-resolution algorithm of single face image was proposed.Based on Markov network model,the study of semantic similarity was introduced.It Limited the scope of the study in the location-related area of feature semantic,and also enhanced the efficiency of learning algorithms and the reconstructed image fidelity.The introduction of Weight fusion mechanism improved the high frequency components of the output image and it made the global effect look better.Analysis and experiments show that the algorithm can learn quickly more valuable image information in large number training-sets.In the other hand,it can suppress effectively image distortion and improved the quality of super-resolution image.
出处
《微计算机信息》
2012年第1期148-150,共3页
Control & Automation
基金
基金申请人:黄东军
项目名称:语义马尔可夫网络下人脸图像非线性超分辨率算法
基金颁发部门:国家自然科学基金委(60873188)
中南大学研究生学位论文创新资助项目(2009ssxt199)
关键词
人脸图像
超分辨率
语义相似度
权值
face image
super-resolution
semantic similarity
weight
