一种适应不同距离的低清人脸深度识别算法  

Deep recognition of low-res faces in varying distances

作　　者：邵文泽[1] 胡洪明 李金叶 邓海松[2] SHAO Wenze;HU Hongming;LI Jinye;DENG Haisong(School of Communications and Information Engineering,Nanjing University of Posts and Telecommunications,Nanjing 210003,China;School of Statistics and Data Science,Nanjing Audit University,Nanjing 211815,China)

机构地区：[1]南京邮电大学通信与信息工程学院,江苏南京210003 [2]南京审计大学统计与数据科学学院,江苏南京211815

出　　处：《南京邮电大学学报（自然科学版）》2023年第1期1-10,共10页Journal of Nanjing University of Posts and Telecommunications：Natural Science Edition

基　　金：国家自然科学基金(61771250,61972213,11901299)资助项目。

摘　　要：针对多数人脸识别算法对于实际低清影像鲁棒性弱的问题,构建了一种融合知识蒸馏和域自适应的低清人脸识别新模型,包含教师网和学生网。首先,两分支骨干网均引入Res2Net模块,以助于提取细粒度强的人脸身份特征;其次,骨干网的不同阶段均引入知识蒸馏,以助于提升学生网的低清人脸特征提取能力;最后,在学生网引入域自适应学习机制,以助于实现域不变的特征提取能力。公开数据集上的实验结果验证了新模型对于不同距离低清人脸的有效性。Given the low robustness of current recognition methods for real low-res faces, this paper designs a novel deep face recognition model via combining knowledge distillation and domain adaptation, including the teacher and the student branches. Firstly, the Res2Net module is borrowed for backbones of the proposed model to extract more fine-grained facial features. Secondly, the knowledge distillation is applied to every stage of the backbones to boost the feature extraction capability of the student net from low-res faces. Finally, the domain adaption is incorporated into the student net to achieve a domain-invariant feature extraction for real low-res faces. Experimental results on a public benchmark dataset demonstrate the effectiveness of the proposed model on low-res faces in varied distances.

关 键 词：人脸识别 知识蒸馏 域自适应 Res2Net 视频监控 

分 类 号：TN911.73[电子电信—通信与信息系统] TP391[电子电信—信息与通信工程]