基于上下文感知和注意机制的多学习情绪识别方法  

Multi learning emotion recognition based on context awareness and attention mechanism

作　　者：万家华[1] 陈乃金 WAN Jiahua;CHEN Naijin(School of Information Engineering,Anhui Xinhua University,230088,Hefei,Anhui,China;School of Computer and Information,Anhui Engineering University,241000,Wuhu,Anhui,China)

机构地区：[1]安徽新华学院信息工程学院,安徽合肥230088 [2]安徽工程大学计算机与信息学院,安徽芜湖241000

出　　处：《北京师范大学学报（自然科学版）》2021年第5期601-605,共5页Journal of Beijing Normal University(Natural Science)

基　　金：国家自然科学基金资助项目(61973295);安徽省教育厅重点科研资助项目(KJ2019A0877)。

摘　　要：为提高人脸图像情绪识别效率与准确性,在探讨了深度神经网络、注意机制与损失函数基础上,提出基于上下文感知与注意机制的多学习情绪识别网络结构.该网络主要由场景特征提取、身体特征提取与融合决策3个子网络组成,并采用单双输出结构,实现多标签情绪分类与连续空间情绪回归任务.考虑到多标签情绪分类时标签的不平衡性,提出了一个改进的焦点损失(focal loss,FL)函数,可为小样本或难分类样本分配更多的权重,从而提高了网络训练效率.利用EMOTIC数据集进行仿真,结果表明平均绝对误差回归组合损失训练性能更优,分类平均准确率与回归平均误差率分别为28.5%和0.098,该方法对于小样本或难分类样本具有更好的分类效果.To improve the efficiency and accuracy of facial image emotion recognition,a multi learning emotion recognition network structure was proposed based on context awareness and attention mechanism. The proposed network was composed of three sub networks: scene feature extraction,body feature extraction and fusion decisionmaking. Single and double output structures were adopted to realize multi-label emotion classification and continuous spatial emotion regression. Improved focus loss function was proposed to assign more weights to small samples or samples that were difficult to classify,to improve efficiency of network training. Simulations using emotic data set showed that proposed improved focus loss and mean absolute error regression combination loss was better,average classification accuracy and regression average error rate were 28.5% and 0.098 respectively. It is concluded that the proposed method had better classification effect for small samples.

关 键 词：人脸图像 情绪识别 上下文感知 注意机制 多标签 

分 类 号：TP393[自动化与计算机技术—计算机应用技术]