文献类型：期刊文献

中文题名：基于卷积神经网络融合SIFT特征的人脸表情识别

英文题名：FACE EXPRESSION RECOGNITION BASED ON CONVOLUTIONAL NEURAL NETWORK FUSING SIFT FEATURES

作者：张俞晴[1];何宁[2];魏润辰[1]

第一作者：张俞晴

机构：[1]北京联合大学北京市信息服务工程重点实验室,北京100101;[2]北京联合大学智慧城市学院,北京100101

第一机构：北京联合大学北京市信息服务工程重点实验室

年份：2019

卷号：36

期号：11

起止页码：161-167

中文期刊名：计算机应用与软件

外文期刊名：Computer Applications and Software

收录：CSTPCD;;北大核心:【北大核心2017】；

基金：国家自然科学基金项目(61572077,61370138,61872042)

语种：中文

中文关键词：卷积神经网络;SIFT特征;视觉词袋模型;特征融合;表情识别

外文关键词：Convolutional neural network;SIFT features;Visual bag of words model;Feature fusion;Expression recognition

摘要：表情识别技术是计算机从静态表情图像或动态表情图像中识别出特定的表情,是实现人机交互的基础.提出一种融合卷积神经网络(CNN)与SIFT特征的人脸表情识别方法.通过图像预处理得到规范化的表情图像;采用视觉词袋模型将图像提取的SIFT特征作进一步处理,将得到的图像特征向量作为局部特征,CNN提取的特征作为全局特征,全局特征用以描述表情的整体差异,局部特征用以描述表情的局部差异;将提取出的两组特征融合后采用Softmax分类.与流形稀疏表示(Manifold Sparse Representation,MSR)及3DCNN等方法在CK+及FER2013数据集上的实验表明,该方法是一种有效的表情识别方法.
The expression recognition technology is that a computer recognizes a specific expression from a static expression image or a dynamic expression image,and is the basis for realizing human-computer interaction.This paper proposed a face expression recognition method that combined convolutional neural network(CNN)and SIFT features.We performed image preprocessing to obtain a normalized expression image.Then the visual bag of words model was used to further process the SIFT features extracted by images,we used image feature vector as a local feature and CNN features as global features.Global features were used to describe the overall difference in expressions.Local features were used to describe local difference in expressions.After the fusion of the two features,Softmax classification was adopted to classify them.Experiments on CK+dataset and FER2013 dataset with manifold sparse representation(MSR)and 3D CNN show that this method is an effective expression recognition method.