Robust facial expression recognition via lightweight reinforcement learning for rehabilitation robotics

作     者：CHEN Yifan FAN Weiming GAO Hongwei YU Jiahui JU Zhaojie 

作者机构：School of Computing University of Portsmouth School of Automation and Electrical Engineering Shenyang Ligong University Department of Biomedical EngineeringZhejiang Univeristy 

出 版 物：《Optoelectronics Letters》 (光电子快报(英文))

年 卷 期：2024年

学科分类：1002[医学-临床医学] 080202[工学-机械电子工程] 08[工学] 1010[医学-医学技术（可授医学、理学学位）] 080203[工学-机械设计及理论] 0804[工学-仪器科学与技术] 0802[工学-机械工程] 100215[医学-康复医学与理疗学] 10[医学] 

基　　金：supported The National Natural Science Foundation of China (52075530) the AiBle project co-financed by the European Regional Development Fund LiaoNing Province Higher Education Innovative Talents Program Support Project (Grant No. LR2019058) Scientific Research Project of Liaoning Education Department (Grant No. LG201909) LiaoNing Province Joint Open Fund for Key Scientific and Technological Innovation Bases(Grant No.2021-KF-12-05) the Zhejiang Provincial Natural Science Foundation of China (LQ23F030001) 

摘      要：The Facial Expression Recognition (FER) capability could endow machines with a more accurate evaluation of rehabilitation robots， making rehabilitation robots’ experience more harmonious and credible. How to obtain robust facial expression representation features has yet to be fully explored. Moreover， the convolutional networks often used in facial expression recognition usually have many parameters， which is not conducive to the online diagnosis of rehabilitation robots. Therefore， this paper proposes a Lightweight Reinforcement Network (LRN) and Auxiliary Label Distribution Learning (ALDL) based robust FER method. Our designed representation reinforcement network mainly comprises two modules， i.e.， the Representation Reinforcement (RR) module and the Auxiliary Label Space Construction (ALSC) module. RR module highlights key feature messaging nodes in feature maps， and ALSC allows multiple labels with different intensities to be linked to one expression. Therefore， LRN has a more robust feature extraction capability when model parameters are greatly reduced， and ALDL is proposed to contribute to the training effect of LRN in the condition of ambiguous training data. We tested our method on FER-Plus， and RAF-DB data sets， and the experiment demonstrates the feasibility of our method in practice during rehabilitation robots.