基于改进损失函数的轻量级人脸表情识别  

Lightweight Facial Expression Recognition Based on Improved Loss Function

作　　者：李娇[1,2] 王彬彬 周心荆 冉峰[1,2] 

机构地区：[1]上海大学微电子研究与开发中心,上海200000 [2]上海大学新型显示技术及应用集成教育部重点实验室,上海200000

出　　处：《工业控制计算机》2021年第6期13-14,17,共3页Industrial Control Computer

摘　　要：针对人脸表情识别网络参数量大、计算资源消耗大的问题,提出了基于改进损失函数的轻量级表情识别算法。改进损失函数先在中心损失函数的基础上增加注意力机制,再融合softmax损失函数共同监督表情识别网络的训练。所设计的网络EERN(Efficient Emotion Recognize Net)仅有0.06M大小,且在RAF-DB数据集上取得了85.3%的准确率。实验证明EERN具有良好的识别性能,有利于模型进一步在嵌入式边缘设备上的部署。At present,many expression recognition neural networks have the problems of large amount of parameters and computing resources.To solve the problem,this paper proposes a lightweight expression recognition algorithm based on improved loss function.The improved loss function first adds attention mechanism on the basis of central loss function,and then combines softmax loss function to supervise the training of facial expression recognition network.Experiments show that the EERN(Efficient Emotion Recognize Net)designed in this paper is only 0.06M,and achieves 85.3%accuracy on RAF-DB dataset.The net designed has good recognition performance,which is conducive to deploy on low hardware resource devices.

关 键 词：表情识别 损失函数 轻量化 特征提取 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]