电子学报2023，Vol.51Issue(10)：2936-2949,14.DOI:10.12263/DZXB.20211352
基于多任务学习和身份约束的生成对抗网络人脸校正识别方法
Multi-task Learning and Identity-constrained Generative Adversarial Network for Face Frontalization and Recognition
摘要
Abstract
For the DR-GAN(Disentangled Representation learning-Generative Adversarial Network),the identity in-formation is not considered in the whole process of generating frontal faces from non-frontal faces with large pose varia-tions.It results in the weak identity consistency between non-frontal faces and the generated frontal faces for disentangling pose from identity.This paper proposes a multi-task learning and identity-constrained generative adversarial network for face frontalization and recognition.Based on the multi-task learning mechanism,a pose classification module and an identi-ty constraint recognition module are constructed between the encoder and decoder of the generative network.These two modules consider the disentangling of face identity and pose in the generating process.More importantly,face identity su-pervision information is added in the process of generating faces from non-frontal faces.In the process of training,identity and pose categories are directly used as the supervision information for learning identity coding features and pose coding features.The identity feature loss function is designed to constrain the identity coding features of the non-frontal faces to approximate the identity coding features of the frontal faces.The effective disentangling of identity and pose information in the non-frontal coding feature is realized.The decoder can more accurately generate a frontal face consistent with the non-frontal face.On the M2FPA dataset,the frontal faces generated from the non-frontal faces with different poses by the pro-posed method are used to recognize,achieving a higher face recognition accuracy.The experimental results show that even when the pose variations are large,the proposed method can still generate a frontal face with a consistent identity,signifi-cantly improving face recognition accuracy under large pose variations.关键词
多任务学习/身份约束/生成对抗网络/人脸校正/人脸识别Key words
multi-task learning/identity constraint/generative adversarial network/face frontalization/face recog-nition分类
信息技术与安全科学引用本文复制引用
黄欣研,刘芳,鲍骞月,李任鹏,刘旭,李玲玲,陈璞花,刘洋..基于多任务学习和身份约束的生成对抗网络人脸校正识别方法[J].电子学报,2023,51(10):2936-2949,14.基金项目
国家自然科学基金(No.62076192) （No.62076192）
国家自然科学基金重点项目(No.61836009) （No.61836009）
长江学者及大学创新研究团队计划(No.IRT_15R53) （No.IRT_15R53）
高等学校学科创新引智计划(No.B07048) （No.B07048）
教育部重点科技创新研究项目 （）
国家重点研发计划 （）
CAAI华为MindSpore开放基金National Natural Science Foundation of China(No.62076192) （No.62076192）
State Key Program of National Natural Science Foundation of China(No.61836009) （No.61836009）
Program for Cheung Kong Scholars and Innovative Research Team in University(No.IRT_15R53) （No.IRT_15R53）
Fund for Foreign Scholars in University Research and Teaching Programs(No.B07048) （No.B07048）
Key Scientific Technological Innovation Research Project by Ministry of Education （）
National Key Research and Development Program of China （）
CAAI Huawei MindSpore Open Fund （）
