摘要
研究了基于学习的人脸图像超分辨率技术。针对Baker方法建立的图像金字塔提取高频细节不够丰富的缺点,提出基于多分辨率幻觉脸算法,采用Kirsch算子提取了高频特征。该算子与Baker的一阶、二阶灰度算子结合,能够提取更多的图像信息,使得匹配更为准确。将流形学习中的LLE算法思想引入匹配复原过程,复原结果获取了更完备的高频信息。对IMDB人脸库进行了试验比较,结果表明,本文方法可取得30.92 dB的平均峰值信噪比,高于Baker方法和插值算法;而且本文预测得到的先验模型更为准确,使得最终复原的人脸图像具有更好的视觉效果。
A novel face hallucination algorithm is proposed based on multi-resolution pyramid struc- ture. In difference from Baker method, the algorithm adds Kitsch operator to get high-frequency features of image. By combined the Kitsch feature with the first and the second order gray features,the new algorithm can extract more image information and make the matching process more accurate. In addition, one of the manifold learning methods, Locally Linear Embedding (LLE) for matching process, is introduced to obtain more high frequency information. The experimental results show that compared with the other methods in IMDB face database, the average PSNR of the proposed algorithm can reach 30.92 dB, which is superior to that of other methods. Moreover, proposed priori model is more accurate, so that the recovery of face image has better visual effect.
出处
《光学精密工程》
EI
CAS
CSCD
北大核心
2008年第5期815-821,共7页
Optics and Precision Engineering
基金
教育部科研基金资助项目(No.107094)
四川大学电子信息学院青年教师基金资助项目
关键词
基于学习的超分辨率
幻觉脸
KIRSCH算子
流形学习
局部线性嵌入
learning-based super-resolution
face hallucination
Kirsch operator
manifold learning
Locally Linear Embedding (LLE)