小波包特征提取及方差相似度的人脸识别

Face recognition using wavelet packet feature extraction and variance similarity

小波包变换是小波变换的推广,可视为普通小波函数的线性组合,具有良好的时频局部性和正交性,随着分解层数的增加,小波包分解能够在所有的频率范围聚焦。利用图像小波包变换的系数矩阵,能够构造出不同的人脸特征向量。针对人脸识别过程中的图像匹配问题,采用计算人脸特征向量方差的方法,并通过方差与权值的对应关系,转换出用于相似度计算的权值。基于理论推导得到的权值具有很好的稳定性,由这些权值计算出的方差相似度也具有较强的适应性,能够减弱由图像噪声、变形等干扰带来的影响。实验表明,该方法识别率高、实时性好。

Wavelet packet transform,which is generalized by wavelet transform,can be considered as a linear combination of general wavelet functions and has good time-frequency local performance and orthogonality.With decomposition level increases,wavelet packet decomposition can focus on range of all frequency.Different facial feature vectors can be constructed by image coefficient matrixes of wavelet packet transform.The image match for face recognition is studied.Variances of feature vectors in relation to facial images are computed,and the weights used for computation of similarity are obtained by a certain transform between the variance and weight.The weights based on the theoretical derivation have good stability.And the variance similarity calculated by these weights has a strong adaptability,weakening the impact of interferences including the noise and deformation of images.The experiments show that the proposed method has the characteristics of high recognition rate and better real-time performance.