基于Contourlet变换的人脸图像超分辨率技术研究  

Face hallucination based on contourlet transform

作　　者：吴炜[1] 杨晓敏[1] 陈默[1] 何小海[1] 郑丽贤[1] 

机构地区：[1]四川大学电子信息学院图像信息研究所,四川成都610064

出　　处：《光电子．激光》2009年第5期694-697,共4页Journal of Optoelectronics·Laser

基　　金：教育部重点项目资助(107094);四川大学电子信息学院青年教师基金的资助

摘　　要：采用Contourlet变换,对基于学习的人脸图像超分辨率技术进行了研究。为了表示人脸特征,首先建立了Contourlet金字塔,针对Contourlet变换特征存在的不足,采用Contourlet变换的方向滤波器(DFB)对高斯金字塔的一阶和二阶梯度特征进行方向滤波,提取滤波后的特征作为新增加的特征。然后针对人脸的特殊性,在匹配过程中采用对应点进行匹配的方法,既保持了匹配的准确性又加快了匹配速度。最后采用最小欧式距离进行最优匹配的选择,对该距离进行判断,针对不同的距离采用不同的复原方法。实验结果表明,本文算法复原出的超分辨率人脸图像具有更好的视觉效果,更逼真、更接近于原始高分辨率图像。This paper presents a new learning based super-resolution of face image by introducing the contourlet transform, which can capture the smoothness along contours by making use of directional decom- position. First,in order to represent the facial features, we establish the Contourlet pyramid. However the transformed feature is not very effective enough. Then,we use DFB to filter the first derivative and second derivative pyramids of the Gaussian pyramid. Finally, from the trained contourlet pyramids, we find the smallest Euclidean distance to the computed feature. And according to the distance, we employ different strategies to deal with different cases. The experimental results show that our method achieve better visual effects and are closer to the original high-resolution images.

关 键 词：基于学习的超分辨率 幻觉脸 CONTOURLET变换 Contourlet金字塔 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]