基于偏最小二乘算法的人脸图像超分辨率技术  

Face Hallucination Based on Partial Least Squares

作　　者：吴炜[1] 杨晓敏[1] 陈默[1] 何小海[1] 

机构地区：[1]四川大学电子信息学院图像信息研究所,成都610064

出　　处：《光子学报》2009年第11期3025-3033,共9页Acta Photonica Sinica

摘　　要：提出了基于偏最小二乘法回归的超分辨率复原算法.介绍了偏最小二乘法回归算法的原理,研究和分析了基于偏最小二乘法回归的超分辨率复原算法.将高低分辨率图像块的高频信息和中频信息作为其特征,并采用分块重叠的方法解决了复原时存在的方块效应.通过对亚洲人脸和欧美人脸的实验结果表明,提出的方法无论是对亚洲人脸还是欧美人脸都能取得较好的复原效果,并且在放大倍数较大的情况下,复原的效果仍然显著.A super-resolution algorithm based on Partial Least Squares （PLS） regression is proposed.PLS regression algorithm is introduced,and a super-resolution algorithm based on PLS regression is analyzed.High resolution image blocks use the high-frequency information as its feature,and low resolution image blocks use middle-frequency as its feature.The block overlapping methods are adopted to solve the block effect of the recovery.The experimental results show that the proposed algorithm performs good to IMDB face database and Yale face database.With 4 fold improvements in resolution （16 times as many pixels）,the algorithm also yield good recovery results.The results of the algorithm are closer to the real images,with a higher peak signal to noise ratio.

关 键 词：基于学习的超分辨率 幻觉脸 偏最小二乘法 回归算法 

分 类 号：TP751[自动化与计算机技术—检测技术与自动化装置]