摘要
为了解决传统的基于人工特征的负面表情识别方法在面部无遮挡、姿态非倾斜的人脸表情图像上表现良好,但是在复杂场景下的识别效果较差的问题,提出了一种基于改进的卷积神经网络的负面表情识别方法.首先利用卷积神经网络的无监督特征学习的特性,预训练两个不同拓扑结构的卷积神经网络,用以提取表情特征;然后融合这些特征,训练分类性能更强的支持向量机.改进后的卷积神经网络算法具有较好的鲁棒性和泛化能力,在训练数据库ICML-fer2013上取得了86.2%的识别率,在测试数据库CK+,GENKI和JAFFE上分别取得了81.6%,87.0%和80.8%的识别率.
The traditional negative facial expression recognition methods based on manual feature perform well on frontal face without facial occlusion,but their performance gets worse in complex condition.To solve this problem,we propose an improved method based on convolutional neural networks(CNN).Firstly,two different architectures of CNN were pre-trained as feature extractors due to CNN′s capability of unsupervised feature learning.The feature CNN extracted was then used to train a more powerful classifier:support vector machine.This improved CNN algorithm has better robustness and generalization,achieving recognition accuracy of 86.2% on the training set ICML-fer2013,and 81.6%,87.0%,80.8% on the testing sets CK+,GENKI,JAFFE respectively.
出处
《华中科技大学学报(自然科学版)》
EI
CAS
CSCD
北大核心
2015年第S1期457-460,共4页
Journal of Huazhong University of Science and Technology(Natural Science Edition)
基金
国家自然科学基金资助项目(60875050
60675025)
关键词
负面表情识别
卷积神经网络
无监督特征学习
特征融合
支持向量机
negative facial expression recognition
convolutional neural network
unsupervised feature learning
feature fusion
support vector machine