Research On Face Recognition Method In Unconstrained Environment

Face recognition is a technology in the field of biometric recognition,and has a large number of application scenarios in security,finance,military,and transportation.The main difficulty of current face recognition technology is that factors in unrestricted scenes such as lighting,pose,and occlusion will seriously affect the stability of the face recognition system.In order to solve this problem,the following researches are proposed in this paper:1.This paper proposes a large-scale face recognition dataset and a test standard for East Asian populations.The 100,000 or higher level face recognition datasets that open sourced on the Internet are mainly aimed at European and American populations,making the usage of the algorithm limited.This paper collects a large amount of facial data of East Asian people.At the same time,this paper uses self-built datasets to make a test standard for testing the effectiveness of face recognition algorithms on East Asian population face recognition.2.This paper studies a method to increase feature distance between different individuals through a cluster center distance constraint.This paper firstly analyzes the spatial distribution of facial features.In order to prevent hard samples between adjacent cluster cores from being misunderstood,this paper uses a constraint term to force each cluster center away from each other during the optimization process.In addition,this paper also presents a simple implementation of the constraint based on the caffe framework.Experimental results verify that the algorithm can significantly increase the heterogeneous feature interval and reduce the probability of face misrecognition.3.This paper studies a face recognition network structure that introduces multi-scale global attention.In this paper,a global spatial attention extraction module is designed based on feature identity transformation,and the module is embedded in three different depth positions on the network.To solve the problem that high-precision face recognition networks are easily affected by noise samples,this paper proposes an adaptive noise filtering method that can automatically filter noise samples during training.Experimental results verify that this method is helpful to improve the accuracy of hard sample recognition.4.This paper studies a feature super-resolution mapping module for low-resolution face recognition.Refers to the idea of knowledge distillation,this paper uses the distillation loss to transfer the detailed information of high-resolution face features as knowledge to the low-resolution face training network,and trains a moudle for mapping lowresolution face features to high-resolution face features.Aiming at solving the problem of training this module,this paper proposes a multi-stage training method to make this module converge.Experiments can prove that the algorithm has a significant improvement effect on low-resolution face recognition.This paper verifies the effectiveness of the above work on LFW,YTF,and self-built test standards.Experimental results show that the algorithms in this paper can effectively improve face recognition accuracy on each standard.