融合Fisher判别分析的多任务深度判别度量学习的化妆人脸验证方法

MAKEUP FACE VERIFICATION METHOD BASED ON MULTI-TASK DEEP DISCRIMINATIVE METRIC LEARNING WITH FISHER DISCRIMINANT ANALYSIS

针对人脸化妆导致人脸验证方法性能降低的问题,提出一种融合Fisher判别分析的多任务深度判别度量学习模型(MT-DDML-FDA)。使用深度度量学习结构,通过共享一个网络层在多个任务之间学习共享的转换知识,来捕获不同任务的人脸图像之间的潜在识别信息;使用Fisher判别分析将类内相关矩阵和类间相关矩阵引入该模型,使每一个任务具有良好的距离度量。实验证明,MT-DDML-FDA在真实化妆人脸数据集上能够有效提升人脸验证的性能。

Face makeup can degrade the performance of face verification methods.To solve this problem,a Learning model of multi-task deep discriminative metric learning with Fisher discriminant analysis(MT-DDML-FDA)is proposed.It used deep discriminative metric learning structure to capture potential recognition information between different tasks by sharing a common network layer to learn projection knowledge among multiple tasks.Meanwhile,it introduced Fisher discriminant analysis,which used the intra-class correlation matrix and inter-class correlation matrix to show a good distance measurement for each task.Experiments show that MT-DDML-FDA can effectively improve the performance of face verification on real makeup face data sets.