摘要： 针对图像集人脸识别中的子空间模型限制,结合Gabor小波变换与K-L高斯黎曼流形判别,提出一种新的图像集人脸识别方法。通过Gabor小波变换表征图像集中人脸图像的特征向量,利用混合高斯模型中带有先验概率的高斯分量表示每个图像数据集,采用可信的K-L概率核函数表示高斯分量间的不同距离,通过加权核判别分析最大化高斯分布间的间距,获取底层数据分布。实验结果表明,与基于线性仿射子空间、基于非线性流形和基于统计模型的方法相比,该方法在YTC和COX数据库上的识别率较高,在YTF数据库上的ROC曲线面积达到85.91,表现最优。测试和训练时间比较结果也表明该方法更适合应用于离线图像集人脸识别系统。

关键词: Gabor小波, 高斯黎曼流形, 人脸识别, K-L核函数, 加权核

Abstract: For the limitations of sub-space model in face recognition with image sets,a face recognition method with image sets based on Gabor wavelet transform and K-L Gaussian Riemannian Manifold(GT-GRMD) discriminant is proposed.Firstly,Gabor wavelet transform is used to characterize feature vectors of faces in image sets.Then,Gaussian components with priori probabilities in Gaussian Mixture Model(GMM) are used to represent each image data set,and credible K-L kernel function is used to represent different distances between Gaussian components.Finally,the proposed weighted kernel discriminant analysis is applied to maximize distance between the Gaussian distributions,accessing the underlying data distribution.Experimental result show that,compared with the method based on linear affinity sub-space,the one based on nonlinear manifold and the one based on statistical model,the proposed method has higher recognition rates on YTC and COX,and the ROC curve area on YTF is up to 85.91,with the best performance.Testing and training time conparision results also show that this method is more suitable for off-line system of face recognition with image sets.

Key words: Gabor wavelet, Gaussian Riemannian manifold, face recognition, K-L kernel function, weighted kernel

中图分类号: 

• TP391.4