基于迁移卷积神经网络的人脸表情识别  

Facial Expression Recognition Based on Migration Convolutional Neural Network

作　　者：刘伦豪杰 王晨辉 卢慧 王家豪 LIU Lun-hao-jie;WANG Chen-hui;LU Hui;WANG Jia-hao(School of Internet of Things Engineering, Hohai University, Changzhou 213022, China;School of Enterprise Management, Hohai University, Changzhou 213022, China)

机构地区：[1]河海大学物联网工程学院,江苏常州213022 [2]河海大学企业管理学院,江苏常州213022

出　　处：《电脑知识与技术》2019年第3期191-194,共4页Computer Knowledge and Technology

基　　金：河海大学常州校区科技基金重点项目

摘　　要：人脸表情识别在计算机视觉领域引起广泛关注,为了解决实际应用中出现的小数据集和硬件限制问题,引入迁移学习方法,将Image-Net上训练好的Inception_v3网络迁移到表情识别任务中,使用FER2013数据集进行参数学习完成表情识别任务,识别率达到了80.4%,且无过拟合现象,网络泛化效果好。基于迁移网络处理的是复杂度更大的分类问题,提取的抽象信息并不都对表情识别任务有利,进一步在迁移学习网络后加入了卷积层和池化层进行表情特征提取和冗余信息筛除,识别率提高到了87.5%。Facial expression recognition has attracted widespread attention in the field of computer vision, and it has played an important role in applications such as human-computer interaction. At the same time, the convolutional neural network method performs well in the graphics recognition task. Therefore, this paper firstly designs a convolutional neural network to complete the expression recognition task. The experimental results show that the recognition rate of the network on the CK+ dataset reaches 99.6%, but the recognition rate on the generalization set is only 21%, showing strong Fitting the phenomenon. In order to avoid over-fitting, the larger fer2013 database is used for training. Under the constraints of hardware conditions, convergence cannot be achieved, and the recognition rate is only 51.7%. In order to solve the small data set and hardware limitation problems in practical applications, this paper introduces the migration learning method, migrates the trained Inception_v3 network on Image-Net to the expression recognition task, modifies the final fully connected layer, and uses the FER2013 data set to perform parameters. After learning to complete the expression recognition task, the experimental results show that the recognition rate reaches 80.4%, and there is no over-fitting phenomenon, and the network generalization effect is good. However, the migration network deals with a more complex classification problem, and the extracted abstract information is not all beneficial to the expression recognition task. Based on this, the migration convolutional neural network designed in this paper adds the convolutional layer and the pooling layer to the expression feature extraction and redundant information screening after the migration learning network. The experimental results show that the recognition rate is increased to 87.5%.

关 键 词：表情识别 卷积神经网络 迁移学习 CK+ FER2013 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]