Facial expression recognition combining attention mechanism and gradient coordination mechanism

• Published in: 中国邮电高校学报（英文版） Vol. 31; no. 6; pp. 35 - 43
• Main Authors: Li Jian, Yang Jun, Li Pengwei
• Format: Journal Article
• Language: English
• Published: School of Information Engineering,Ningxia University,Yinchuan 750021,China 01.12.2024; School of Cyberspace Security,Beijing University of Posts and Telecommunications,Beijing 100876,China%School of Information Engineering,Ningxia University,Yinchuan 750021,China%School of Artificial Intelligence,Beijing University of Posts and Telecommunications,Beijing 100876,China
• Subjects: attention mechanism; facial expression recognition; gradient coordination mechanism; deep effective network
• ISSN: 1005-8885
• DOI: 10.19682/j.cnki.1005-8885.2024.1019

Facial expression recognition poses a formidable challenge due to the nuanced variations in facial expressions,which complicate the extraction of discriminative features.Furthermore,the imbalance in sample category counts exacerbates this intricacy.To address this challenge,an innovative facial expression recognition approach that integrates a cascade framework with an attention mechanism and a gradient coordination mechanism is proposed.Firstly,the attention mechanism is introduced into the deep effective network,and the attention weight is inferred according to the channel and spatial dimension.This mechanism enhances the expressive ability of expression features and suppresses the impact of redundant information.Secondly,the gradient coordination mechanism is used to suppress the contribution of easy-to-sort abnormal samples to the model,and more attention is paid to the contribution of difficulty to classify samples.This approach weakens the impact of the unbalanced distribution of the dataset,thereby improving the recognition rate of the model.Experimental results show that the method can achieve recognition accuracy rates of 73.28％ and 97.14％ on the public datasets FER2013 and JAFFE,respectively.Compared with other algorithms,the proposed method exhibits a certain degree of advancement and superiority.These results illustrate the effectiveness and applicability of the proposed method.