基于改进ORB特征的多姿态人脸识别

Multi-pose Face Recognition Based on Improved ORB Feature

为了克服通过全局特征以及每位个体单个模板样本进行多姿态人脸识别的不足,提出基于改进的ORB局部特征以及每位个体多个模板样本的多姿态人脸识别方法.首先改进ORB算子的采样模式提高算子对人脸视角变化的鲁棒性,并采用每位个体的多个训练样本建立模板库,然后提取并匹配测试样本与模板样本的改进ORB特征.在特征提取阶段,为避免关键点数目变化的干扰,对全部样本提取一致数目的关键点;在特征匹配阶段,采用基于模型和基于方向的双重策略剔除误匹配点对,使用匹配点对数目与平均距离评价测试样本与每个模板样本的吻合程度.对CAS-PEAL-R1和XJTU数据库的实验结果表明,改进的ORB特征具有更好的识别性能;与采用多个训练样本构建个体单个模板样本的方法相比,在训练样本数目相同的条件下,该方法能较好地避免姿态的干扰,具有更好的识别效果.与SIFT算子相比,ORB算子在特征提取与特征匹配2个阶段都具有明显的速度优势.

To overcome the weakness of multi-pose face recognition by global feature or by single template samplefor each subject, a novel multi-pose face recognition method based on improved oriented FAST and rotatedBRIEF （ORB） local feature and multiple template samples for each subject is proposed in this paper. The samplingpattern of ORB operator is first improved to enhance the robustness of the operator to the variation ofviewpoint towards face, and template database is built with multiple training samples of each subject, the improvedORB features of the test sample are next extracted and matched with those of template samples. At thestage of feature extraction, consistent number of keypoints are extracted for all samples to avoid the disturbanceof the variation of the number of keypoints. At the stage of feature matching, double strategies based on the modeland orientation of matching-point pairs are adopted to eliminate outliers. The consistent degree of test sample andeach template sample is evaluated by the number and average distance of the inliers. Experimental results on theCAS-PEAL-R1 and XJTU databases show that, the improved ORB operator has better recognition performance;compared with the methods of constructing single template sample from multiple training samples for each subject,the proposed method could better avoid the disturbance of pose variation, and obtain better recognition resultsunder the condition of using the same number of training samples. Compared with scale invariant feature transform （SIFT） operator, the ORB operator has obvious advantages in speed at both stages of feature extraction and feature matching.