基于多模板卷积的人脸确认算法  

Algorithm for Face Determination Based on Multi-templates and Convolution

作　　者：蓝章礼[1] 王丹妹[1,2] 兰全祥 

机构地区：[1]重庆交通大学信息科学与工程学院,重庆400074 [2]承德医学院生物医学工程系,河北承德067000 [3]攀枝花学院数学与计算机学院,四川攀枝花617000

出　　处：《小型微型计算机系统》2016年第5期1048-1051,共4页Journal of Chinese Computer Systems

摘　　要：提出一种多模板卷积与动态规划相结合的人脸确认算法,利用阈值原理判断待检图像是否为人脸.首先利用人脸样本图像训练得到人脸模板、双眼模板和嘴鼻模板的权重系数和最优阈值;然后裁剪出待检图像人脸、双眼、嘴鼻区域并进行尺寸、像素值归一化处理;之后计算不同区域与其模板卷积值,并利用模板权重加权求和得到人脸确认最终卷积值;最后通过最优阈值与卷积值关系判断待检图像是否为人脸.为验证方法的正确性、有效性,分别利用人脸图像、非人脸图像进行实验,在本文使用的实验库中成功确认率达94%、84%以上.A face determination algorithm based on multi-template is proposed, which is combined with convolution and dynamic pro- gramming. Firstly, three kinds of templates are synthesized using the experimental face database, including face, eyes, and mouth-nose templates, and then obtaining their weighting factors and the optimal threshold. Secondly, the area of face, eyes, mouth and nose is cut from the image, then realizing the size and number normalization, and calculating the three convolution values for detected image and the templates, and then the sum convolution values for face detecting are got by a weighted summation of them. Finally, according to the principle of threshold decision to judge whether the image is human face or non-face. Experimental verifications are conducted to evaluate the algorithm＇s performance on the acquisition, human face and non-face databases. The results show that the algorithm runs faster,the right rates are 94% ,84%.

关 键 词：人脸检测 卷积 多模板 阈值决策 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]