大样本库人脸识别的分级弹性匹配算法

Classified Elastic Matching Face Recognition Algorithm Used in Large Database

提出了一种提高弹性匹配人脸识别算法速度的新算法。弹性匹配具有较高的识别率,但计算复杂度较高,影响了其在大样本库中的应用。为此提出分级弹性匹配:将弹性匹配的两个步骤(网格平移和网格变形)中的网格平移看作独立的匹配算法;对所有备选人脸图像先做网格平移计算出粗匹配度(CMS);根据CMS值将所有人脸图像降序排列,只对CMS值较高的部分图像做网格变形。在100180人的人脸图像库上的测试结果表明:相对于传统的弹性匹配,分级弹性匹配算法能在识别率的损失不大于0.5%的前提下,将网格变形的计算量降低1000倍或者更多。

An algorithm to speed up elastic matching face recognition algorithm was presented. Elastic matching has relatively higher recognition rate but also higher calculation complexity, which limits its application. The main idea of Classified Elastic Matching (CEM) is as follows. Regard grid translation, which is one of the two steps of elastic matching, as an independent matching algorithm. To all candidate face images, do grid translation and calculate corresponding Coarse Matching Score (CMS) at first. Align the candidates in descending order according to CMS. Only to grid distortions to these face images having relatively higher CMS values. Experimental results based on a large database including face images from 100180 individuals show that the time cost of grid translation in CEM is more than 1000 times reduced compared to that in traditional elastic matching with recognition rate depressed less than 0.5%.