基于改进的LBP人脸识别算法

The Face Recognition Algorithm Based on Improved LBP

针对基本LBP算子提取的特征不够完整,不能全面地表达出人脸局部特征的问题,提出了基于分块的完备局部二值模式(CLBP)人脸识别算法。首先对原始人脸图像进行分块处理,对每一分块的图像进行局部差异值和中心像素灰度值分析,用Su2CLBP(8,2)、Mu2CLBP(8,2)和CCLBP(8,2)算子分别提取每一分块的直方图统计特征。然后将所有分块的CLBP直方图序列连接起来,得到人脸图像的CLBP特征,将其作为人脸的鉴别特征用于分类识别。最后利用Chi平方统计法计算直方图的不相似度,用最近邻准则进行分类。所提出的算法分别在ORL、FERET、YALE数据库中进行实验,分别取得了高达99.5%、92%、98.67%的识别率,与分块LBP算法相比识别率分别有2.5%、8%、2.67%的提高。实验结果表明,完备LBP提取的特征比较全面且具有较强的鉴别能力,在ORL、FERET、YALE人脸库中均能获得较好的识别率。

Since the feature extracted by the basic LBP operator are not complete and can not fully represent the local feature of face, a face recognition algorithm based on block completed local binary pattern is proposed. Firstly, the original face image is divided into small blocks from which the local difference value and central pixel grayscale value are analyzed. Extracting the historgram statistical characteristics of each block by the S^u2 CLBP （8,2）. MS^u2 CLBP （8,2） and CCLBP（8,2） operator. Then, the CLBP histograms of all the blocks are linked to get the CLBP feature to be used as the face descriptor. Finally, the classification is performed using a nearest neighbor classifier with Chi square as a dissimilarity measure. Experimental results on ORL. FERET face database show that the proposed algorithm can achieve high face recognition rate up to 99.5%, 92% and 98.67%, which are 2.5%, 8% and 2.67% higher than the block LBP algorithm. This work demonstrates that the completed LBP feature is complete and highly discriminable and has good performance in the ORL, FERET and YALE face database.