轮廓波变换的面部姿态检索技术及实验分析

Facial Profile Retrieval Based on Contour Wavelet Transform and Relevant Experiment

基于经典的面部姿态识别技术进行了人脸识别算法改进。根据LIOP和LBP的优势和问题,提出领域差异向量的方法(YP)实现聚类算法改进,采用了16个领域点的码本,为了不增加计算量,又将图像划分了36个子块;基于传统的PCA算法,提出WPCA实现降维,采用了在PCA的基础上再乘以权重的办法,提出了余弦相似度的判别方法;基于PP和retina模型方法,提出了PR方法和预处理流程。通过中科院的CAS的人脸数据平台进行了算法对比实验:该改进方法在遮挡、背景、年龄、表情以及距离5种测试集中,在鉴别能力,提取时间方面都有了显著改进;与传统算法相比也有了明显的改进,比LBP、PP算法提高了20%以上,有一定的异质人脸识别功能。

Based on the classical facial gesture recognition technology,the face recognition algorithm is improved. Based on advantages and problems of LIOP and LBP,the clustering algorithm developed by area difference vector method（ YP） is improved. 16 points in the field are coded,in order not to increase the amount of calculation,each image is divided into 36 subblocks. Based on the traditional PCA algorithm,WPCA is proposed to reduce the dimension. By PCA and multiplied weighting method,cosine similarity discriminating method is proposed. Based on PP and retina model method,this paper puts forward that PR method and pretreatment process. By face data platform of CAS issued by the Chinese Academy of Sciences,comparing experiment are carried out. The design method has improvements in shade,background,age,facial expression and the distance,the ability of identification,extracting time are all significantly improved. In heterogeneous face recognition,the proposed method has obvious advantage comparing with the traditional algorithms,it is more than 20% higher than that of LBP and PP algorithms,and has the certain heterogeneous face recognition.