基于 PCA和 LCN单样本光照不变人脸识别研究

摘要

Traditional illumination invariant face recognition for single training sample based on principal component analysis ( PCA) and local contrast normalisation ( LCN) uses histogram equalisation and local contrast normalisation to eliminate the effect of uneven illumination, employs Mahalanobis distance to measure the distance between data, and uses mean filter to calculate every pixels neighbourhood expectation in image.On pose illumination expression ( PIE) and extended Yale B face databases we randomly extract the single training sample to evaluate different illumination preprocess methods, distance measures, and two calculation methods for pixels neighbourhood expectation. Experimental results show that to preprocess the illumination with global and local contrast normalisation, to use simplified Mahalanobis distance for distance metric, and to calculate the pixels neighbourhood expectation with two-dimensional Gaussian filter can achieve best recognition results.%传统的基于PCA（ Principal Component Analysis）和局部对比度正则化的单样本光照不变人脸识别，使用直方图均衡化和局部对比度正则化削弱不均匀光照的影响，使用马氏距离度量数据间的距离，使用均值滤波器计算图像各像素邻域的期望。在PIE （ Pose Illumination Expression）和Extended Yale B人脸数据库上，随机抽取单训练样本评估多种光照预处理方法、距离度量和两种计算像素邻域期望的方法。实验结果表明，使用局部和全局对比度正则化进行光照预处理，距离度量使用简化的马氏距离，使用二维高斯滤波器计算像素邻域的期望可以达到最好的识别效果。