Robust Algorithms for Real-World Face Recognition

一个实用的人脸验证系统有望达到0.01%或更低的错误接受率，同时确保错误拒绝率(FRR)为5%或更低。近年来，基于深度神经网络的人脸识别算法在人脸识别库上的测试取得了较高的准确率。然而，当这些算法投入实际应用时，由于光照条件、摄像头距离、人脸方向和遮挡等因素，这些算法往往不够健壮。作为一家初创公司，Altom View Systems Inc.自2015年1月以来一直在开发各种智能摄像头产品。在这个项目中，大学团队将与AlumView合作，提高其现有人脸识别产品的性能和健壮性。将进行的研究包括：1)尝试深度神经网络方案的代价函数的不同配置，并确定最有效的方法；2)设计新的神经网络架构，该架构对ID照片和狂野输入照片之间不同的光照条件、摄像头距离和人脸角度等具有鲁棒性；3)设计方案来处理由眼镜等引起的部分遮挡人脸；4)研究生成对抗网络用于输入照片和域自适应的预处理；以及5)研究如何结合不同的技术来提高系统的整体性能。

A practical face verification system is expected to achieve a False Accept Rate of 0.01% or lower whileensuring a False Reject Rate (FRR) of 5% or lower. In recent years, face recognition algorithms based on deepneural networks have achieved high accuracy when tested on face recognition database. However, when putinto real-world application, those algorithms are often not robust enough, due to factors such as differentlighting conditions, camera distances, face orientations, and occlusions.As a startup company, AltumView Systems Inc. has been developing various intelligent camera products sinceJanuary 2015. In this project, the university team will work with AltumView to improve the performance androbustness of its existing face recognition products. Research to be conducted includes: 1) try differentconfigurations of the cost functions of the deep neural network scheme, and identify the most effective way; 2)design new neural network architecture that is robust to different lighting conditions, camera distances, andface angles etc. between the ID photos and wild input photos; 3) design schemes to deal with partially occludedfaces caused by e.g., glasses; 4) investigate Generative Adversarial Networks for preprocessing of input photosand domain adaptation; and 5) investigate how to combine different techniques to improve the overallperformance of the system.