对称性的图像梯度方向在人脸识别中的应用  

Application of Image Gradient Orientations of Symmetry in Face Recognition

作　　者：段练 王梨清 DUAN Lian;WANG Liqing(Medical School, Nantong University, Nantong, Jiangsu 226019, China;Modern Educational Technology Center, Nantong University, Nantong, Jiangsu 226019, China)

机构地区：[1]南通大学医学院,江苏南通226019 [2]南通大学现代教育技术中心,江苏南通226019

出　　处：《宜宾学院学报》2018年第12期34-37,50,共5页Journal of Yibin University

摘　　要：人脸识别中,样本图像很容易受到光照、表情以及姿态等这些外部因素变化的影响,通过利用图像梯度方向代替传统的像素强度来表示像素之间的相关性,一定程度上缓解了这个问题.然而训练样本的有限性很难准确地描述原始样本图像产生的变化,特征提取过程中容易导致样本图像部分信息的损失.利用原始样本的镜像原理重构对称样本图像,在图像梯度方向的基础上,提出对称性的图像梯度方向人脸识别方法(S-IGO),将S-IGO方法分别与PCA、LDA、IGO-PCA、IGO-LDA以及扩展的E-IGO方法在不同人脸库上的识别结果进行比较,并分析改进算法的在人脸识别中的优势.实验结果表明,相较PCA、LDA和IGO方法,E-IGO方法和S-IGO方法通过利用样本图像的镜像原理生成对称样本图像,在拓展样本集合上进行特征提取,可以选择更稳定的特征空间,从而提高最终的识别结果.大部分情况下, S-IGO方法的识别结果要优于E-IGO方法,这是因为S-IGO算法在E-IGO方法的基础上,进一步利用了人脸对称性这个先验信息,在降维过程中,尽可能多地保留了原始样本的有效信息,提高了算法的准确度.In face recognition,sample image is easily affected by the illumination,facial expression and posture change.To some extent,this problem is mitigated by using image gradient orientations to replace traditional pixel intensities to represent the correlation between pixels.However,the limitation of training samples cannot comprehensively convey the changes of original sample images,and partial information loss of sample images is easily caused in feature extraction process.Symmetric sample image was reconstructed by using image principle of original samples.Then S-IGO method was proposed based on the image gradient direction of symmetry.Finally,the S-IGO method was compared with the recognition results of PCA,LDA,IGO-PCA,IGO-LDA and extended E-IGO methods in different face databases,and the advantages of the improved algorithm in face recognition were analyzed.The experimental results show that the E-IGO and S-IGO methods can generate symmetric sample images by using image principle of original samples,to select more stable feature space and get better recognition results which are extracted from extended sample sets.But in most cases,S-IGO method is superior to E-IGO method with better accuracy because the former can retain the effective information of original samples as much as possible by using face symmetry in dimension reduction process.

关 键 词：图像梯度方向 对称性 人脸识别 线性鉴别分析 

分 类 号：TP312[自动化与计算机技术—计算机软件与理论]