运用竞争反样本的人脸认证算法

Face verification using competitive negative samples

为了在认证中增强类间可鉴别性,通过引入竞争反样本,提出一种新的人脸认证算法。算法中,测试人脸不仅与所声明客户人脸进行匹配比对,同时也与各竞争反样本逐一计算匹配分值。所有分值综合起来,形成最终认证决策。设计最近邻反样本决策、全体反样本决策、最近邻域反样本决策等3种方案,并结合开集模式的人脸认证领域分别在多个人脸库、不同特征和不同分类器上进行实验和比较。在该文的测试中,新算法的3种方案与原有基于相似度的认证算法相比,错误率依次平均降低25.13%、30.24%、30.97%。

A novel verification algorithm was proposed using competitive negative samples to enhance discrimination in face verification. In the algorithm, the test face was matched not only with the claimed client face, but also with competitive negative samples, with all the matching scores combined for a final decision. Three schemes were designed. They were the closest-negative- sample scheme, the all-negative-sample scheme, and the closest- few-negative-sample scheme. The schemes were compared with the traditional similarity-based verification approach on several databases, features and classifiers. The tests demonstrat that the three schemes reduced the verification error rate by 25.13%, 30.24% and 30.97% on average.