摘要
基于2DGabor变换的人脸特征描述已经受到了很多人的关注。然而现有的Gabor特征维数较高,而且具有冗余性,因此选择最佳的Gabor特征用于人脸识别显得尤为的重要。利用最大余量原理的特征选择算法在目前的机器学习研究中已经占据了重要的地位。本文在基于余量的迭代搜索法(Simba)的基础上,引入了一种新的选择算法:基于余量的共轭梯度法(Cgmba),它只需较少次迭代就可以找到最佳解。我们在IMM人脸库上进行了实验,实验结果表明:尽管只使用了一半不到的特征,但Cgmba和Simba的识别率却分别提高了3.75和1.25个百分点,同时也证实了我们提出的Cgmba明显优于Simba。最后我们对Cgmba选择的Gabor特征的分布情况进行了分析,可以看出较大尺度的特征相对于较小尺度的特征对于分辩人脸的细微差别具有同等的重要性,而且在垂直,135°方向的特征具有更强的分辩能力。
Face representation based on 2D Gabor has attracted much attention. However, due to the fact that Gabor features currently are redundant and too high dimensional, selection of optimal Gabor features for face recognition appears to be paramount. Margin-based algorithms which use the large margin principle for feature selection have already played a crucial role in current machine learning research. In this paper, based on iterative search margin-based algorithm （Simba）, we introduce a new selection algorithm： Conjugated gradient margin-based algorithm （Cgmba）, which can find optimal solution at less iteration. Experiments were carried out on IMM face database. Results indicate that Cgmba and Simba can provide 3.75, 1.25 percent improvement in classification rate respectively, though less than half of all features are used. Moreover, superiority of our proposed approach to Simba is also demonstrated. Finally, the distribution of Gabor features selected by Cgmba is analyzed. It is inferred that the features in the larger scales have the same importance as those in the smaller scales in discriminating nuance of faces and features in vertical, and 135° orientations have more discriminative power.
出处
《光电工程》
EI
CAS
CSCD
北大核心
2006年第9期85-90,共6页
Opto-Electronic Engineering
