一种改进的基于小波变换的人脸光照归一化方法

An Improved Wavelet-Based Illumination Normalization Algorithm for Face Recognition

针对光照对人脸识别影响的问题,提出了一种改进的基于对数域多级小波分解的光照归一化方法。首先将手工裁剪后的人脸图片进行对数变换,然后将对数域图像进行多级小波分解,将低频系数置零,并将多个尺度上的高频系数分别乘以不同的高频增益,突出光照不变性分量。采用经典的PCA人脸识别算法,在耶鲁B与CMU PIE人脸数据库的实验结果表明,本文方法能有效地消除光照对人脸识别的影响,并有效提高识别率。

To counter the problem of lighting effect on face recognition, the authors proposed an improved wavelet-based illumination normalization algorithm. Firstly, cropped raw face image was transformed from spatial domain to logarithm domain using logarithm transform algorithm, and then multilevel wavelet decomposition was applied on logarithm domain image. The coefficients of low frequency subbands were set to zero and the coefficients of high frequency subbands were multiplied by different gains in different scales to emphasize illumination invariant. Well-known PCA- based face recognition method was used as classifier and the algorithm was tested in Yale B database and CMU PIE database. The experimental results showed that the proposed method could reduce the influence of illumination and improve recognition rate efficiently.