文献类型：期刊文献

中文题名：多维连续空间的多任务表情识别研究

英文题名：Multi-task Facial Expression Recognition Based on Multi-dimensional and Continuous Space

作者：霍奕[1]

第一作者：霍奕

机构：[1]北京联合大学师范学院教育信息技术系,北京100011

第一机构：北京联合大学师范学院教育信息技术系|北京联合大学特殊教育学院信息技术系

年份：2024

卷号：23

期号：5

起止页码：17-23

中文期刊名：软件导刊

外文期刊名：Software Guide

基金：教育部人文社会科学研究项目(23YJE880001)。

语种：中文

中文关键词：多维情感识别;多任务学习;VAD面部表情识别数据集;离散情感类别;智能情感交互

外文关键词：multi-dimensional emotion recognition;multi-task learning;VAD facial expression recognition dataset;discrete categorial emotion recognition;intelligent emotion recognition

摘要：研究设计一个多任务情感识别模型,通过结合Valence、Arousal、Dominance(VAD)三维连续情感分析与离散情感分类,为智能情感交互提供更全面、细致的情感测量工具。首先利用两个识别任务间的相关约束(类别标签为VAD三维情感空间中的点)提升模型的识别准确性;其次提供一种在VAD三维空间中识别多维连续情感的方法与数据集,利用他们之间的相关性进行多任务联合学习,并在情感类别和VAD多维情感空间之间建立约束,相较于传统固定情感类别标签能更全面、细致地描述情感状态,特别是在目前较少研究的维度D上;最后使用情感类别数据集FER2013中可用的情感标签与手动添加的VAD注释测量VAD情感。实验表明,V和类别、A和类别、D和类别的多任务学习能明显改善模型的识别性能。
A multi task emotion recognition model was designed to provide a more comprehensive and detailed emotion measurement tool for intelligent emotion interaction by combining Valence,Arousal,Dominance(VAD)three-dimensional continuous emotion analysis and discrete emotion classification.Firstly,utilize the relevant constraints between two recognition tasks(labeled as points in the VAD three-dimensional emotional space)to improve the recognition accuracy of the model;Then,a method and dataset for identifying multi-dimensional continuous emotions in VAD three-dimensional space were provided,utilizing their correlation for multi task joint learning,and establishing constraints between emotion categories and VAD multidimensional emotion space.Compared with traditional fixed emotion category labels,it can more comprehensively and meticulously describe emotional states,especially in dimension D,which is currently less studied;Finally,use the sentiment labels available in the sentiment category dataset FER2013 and manually added VAD annotations to measure VAD sentiment.The experiment shows that multi task learning with V and category,A and category,and D and category can significantly improve the recognition performance of the model.