分数次幂多项式核函数在核直接判别式分析中的应用

Application of fractional power polynomial kernel function to kernel direct discriminant analysis

提出了在核直接判别式分析(KDDA)中采用分数次幂多项式核函数的方法,并在ORL人脸库中对多头部姿态、尺度等变化进行了实验。实验结果表明,采用分数次幂多项式核函数比采用整数次幂多项式核函数时的KDDA误识别率明显要低(取36个特征数时,误识别率低2%),且随着使用的特征数不断减少,这种优势愈加明显。实验充分证实了在KDDA中采用分数次幂多项式核函数的有效性及其对人脸的光照、头部姿态、面部表情等变化的鲁棒性。

A method of Kernel Direct Discriminant Analysis （KDDA） using fractional power polynomial kernel function is proposed,and the experiments are carried out for the variations of multi head pose and multi scale in ORL face database. The experiment results indicate that the KDDA misrecognition using fractional power polynomial kernel function is distinctly lower than that using integer power polynomial kernel function（lowers 20/4o when 36 features is selected）. Furthermore, this advantage becomes more evident with the decreasing of the feature number. The experiments fully demonstrate the validity of KDDA using fractional power polynomial kernel function and the robustness to face illumination,head pose,facial expression and other variations.