具有动态调节功能的Fisherface方法

Fisherface Method with Dynamic Adjusting Function

高维、小样本数据的识别问题,始终是模式识别领域的热点和难点问题。由于训练样本数量很少,当以样本的协方差矩阵作为模式协方差矩阵的估计时,会产生较大的偏差。这是造成模式分类错误的主要原因。本文在详细论述 Fisherface 方法的基础上,提出了具有动态调节功能的 Fisherface(DA-Fisherface)方法。该方法利用测试样本完成了对样本协方差矩阵的动态调节,减小了因样本数目很少所造成的偏差,从而实现了对 Fisher 鉴别矢量集的优化。最后,在 ORL 人脸库上的实验结果表明,该方法的模式分类正确率比 Fisherface 方法有显著提高。

In pattern recognition, the classification of high-dimensional and limited-sample data is not only a hotspot but also a difficulty all the time. Because the number of training samples is very small, big bias will occur when pattern covariance matrixes are estimated by training Sample covariance matrixes. It is also an important reason of wrong classification. On the basis of explaining Fisherface method in detail, this paper proposes a Fisherface method with dynamicadjusting function （DA-Fisherface）. The method completes dynamic adjusting the sample covariance matrixes using testing samples, reduces the bias caused by limited training samples, and optimizes the Fisher discriminant vectors. Finally, the experimental results on ORL face database indicate the purposed method gets a higher recognition ratio than Fisher face.