基于改进Kaze特征匹配的低分辨人脸识别算法  

Low Resolution Face Recognition Based on Improved Kaze Feature Matching Algorithm

作　　者：傅敏[1] 刘本永[2,3] 

机构地区：[1]贵州大学计算机科学与技术学院,贵州贵阳550025 [2]贵州大学大数据与信息工程学院,贵州贵阳550025 [3]贵州大学智能信息处理研究所,贵州贵阳550025

出　　处：《贵州大学学报（自然科学版）》2016年第3期102-105,共4页Journal of Guizhou University:Natural Sciences

基　　金：科技部国际合作项目(2009DFR10530)

摘　　要：针对普通算法难以提取低分辨人脸图像特征以实现人脸识别问题,探讨一种基于改进Kaze特征匹配的识别算法。首先,通过改进P-M扩散滤波中阈值和扩散函数来保留低分辨图像的边缘和细节,并利用加性算子分裂算法构造非线性尺度空间;其次,通过寻找不同尺度归一化后的Hessian矩阵局部极大值点来检测特征点,并使用M-SURF构造特征描述向量;最后,利用欧式距离进行特征向量的匹配实现识别分类。实验结果表明,与基于SIFT、SURF和普通Kaze特征匹配的算法相比,所探讨算法正确识别率更高,同时对噪声图像也有更好的鲁棒性。The general algorithm has difficulty in extracting the low resolution face image feature,a recognition algorithm based on improved Kaze feature matching was explore to solve face recognition problem. First,low resolution image edges and details can be saved by improving threshold and diffusion function of P-M diffusion filtering and a nonlinear scale space was constructed by additive operator splitting algorithm. Then the feature points were detected by searching local maximum points of Hessian matrix by different scales normalization and the feature descriptors were constructed by M-SURF. At last,the recognition and classification were relied on matching of feature descriptors by Euclidean distance. Experiments show that the discussed algorithm has better rate of recognition comparing to feature matching algorithm based on SIFT,SURF and general Kaze. Meanwhile,the proposed algorithm has better robustness on noisy image.

关 键 词：低分辨人脸识别 P-M扩散滤波 Kaze特征 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]