开集人脸识别中的性能评估系统

摘要

构建一套适用于开集人脸识别任务的性能评估系统。该系统包括训练集质量评估和对测试样本分类结果性能评估两部分。对于前者，利用巴氏距离作为贝叶斯分类错误率的近似，并考虑开集问题中样本并非服从独立同分布假设的特殊性质，在高斯与非高斯情况下均得到训练集的质量评价函数，其中高斯情况下该函数具有闭式表达。对于后者，通过考察测试样本对附近正/负样本对的分布密度，度量由分类器得到的样本对相似程度的可靠程度，完善以往文献对比缺乏衡量的不足。文中结果在多个人脸数据库上均得到验证。%A performance evaluation system is built for the open set face recognition task. The system consists of two parts:the quality evaluation for training set and the performance evaluation for classification result of test samples. For the former, Bhattacharyya distance is used to approximate the Bayesian error rate, and the particularity of open set problems that sample pairs do not obey the independent identical distribution assumptions is taken into account. The quality evaluation functions of the training set are obtained in both Gaussian or non-Gaussian distribution assumptions, and in Gaussian case this function has a closed form. For the latter, the distribution densities of the nearby positive and negative sample pairs are considered to measure the reliability of the similarity score given by a classifier. Therefore, the previous studies which are lacking of such measurements are complemented. The results in this paper are validated by experiments on multiple face databases.