摘要
针对面部表情识别中Softmax Loss损失函数类内特征分散、类间特征未分离的问题,提出一种改进型Softmax Loss损失函数ACAM Loss。上述损失函数将优化集中在余弦角度上,在特征与目标权重的余弦夹角加上间隔m,在特征与非目标权重的余弦夹角减去间隔m,有效引导网络学习具备类内距离较小、类间距离较大的强区分度特征,同时加大网络训练力度优化输出,以提升分类效果。实验表明,在CK+和JAFFE数据集上,ACAM Loss函数训练模型表情识别准确率分别达到98.87%和98.92%,相比Softmax Loss函数平均增幅分别为1.35%和1.3%,识别准确率处于主流领先水平,验证了改进损失函数的有效性。
To address the problem that the intra-class features are scattered and the inter-class features are not separated in the Softmax Loss loss function in facial expression recognition,an improved Softmax Loss loss function ACAM Loss is proposed.The loss function concentrates the optimization on the cosine angle by adding the interval m to the cosine angle of the feature and the target weight,and subtracting the interval m from the cosine angle of the feature and the non-target weight.The network is effectively guided to learn strong distinguishing features with small intra-class distance and large inter-class distance,while the network training is increased to optimize the output in order to improve the classification effect.The experiments show that on the CK+and JAFFE datasets,the expression recognition accuracy of the ACAM Loss function training model reaches 98.87%and 98.92%,respectively,com-pared with the average increase of 1.35%and 1.3%for the Softmax Loss function,and the recognition accuracy is in the mainstream leading level,which verifies the effectiveness of the improved loss function.
作者
李校林
张鹏飞
董玉晖
钮海涛
LI Xiao-lin;ZHANG Peng-fei;DONG Yu-hui;NIU Hai-tao(School of Communication and Information Engineering,Chongqing University of Posts and Telecommunications,Chongqing 400065,China;Research Center of New Telecommunication Technology,Chongqing University of Posts and Telecommunications,Chongqing 400065,China)
出处
《计算机仿真》
北大核心
2023年第5期234-238,256,共6页
Computer Simulation
关键词
面部表情识别
分类
损失函数
加性余弦角度间隔
Facial expression recognition
Classification
Loss function
Additive cosine angle interval
