特征块权重化的面部表情识别研究  

Facial Expression Recognition Research Based on Feature Block Weighting

作　　者：江渊广 张鹏[1] 王栋 张二东[1] 战茜 

机构地区：[1]上海大学机电工程与自动化学院,上海200072

出　　处：《工业控制计算机》2018年第2期94-95,98,共3页Industrial Control Computer

摘　　要：立足于提高表情识别的正确率,对提取的特征进行分块权重化处理。首先提取面部表情的Gabor纹理特征以及几何特征,并采用PCA降低Gabor特征的维度。针对面部关键轮廓点组成的几何特征,提出将整个面部的几何特征分成左眼、右眼、嘴巴三个特征块,分别进行Procrustes Analysis,将各个特征块进行对齐,解决不同人面部器官大小、位置不一致的问题。针对不同特征、不同特征区域对表情识别贡献率不同的问题,提出基于特征块权重化的Bp神经网络(Feature Block Weighting Bp,FBWBp),在Bp神经网络的输入层前增加了一层权重层,实现对各个特征块的权重化,将特征块的权重化与分类器结合起来,通过训练样本的训练,权重层的权重值和Bp神经网络各层的权重值同时进行搜索优化。选择JAFFE表情库进行表情分类实验验证,实验结果表明该方法可以提高准确率2%到4%左右,该实验提出方法可以提高面部表情的识别正确率。This experiment aims at improving the accuracy of expression recognition and using block weighting treatment.First,the Gabor feature and geometry feature of facial expressions are extracted and PCA is used to reduce the dimension of the Gabor feature.Then,this paper based on the geometry features made up of key facial contour points,proposes to divide the geometry feature of the whole face into three feature blocks,namely left eye,right eye and mouth,which are aligned through Procrustes Analysis.By combining the weighting and classifier of the feature block and through sample training,the weighting value of the weighting layer and of various layers of the Bp nerve net are searched and optimized simultaneously.At last,JAFFE expression database is chosen to conduct expression classification experiment,the result of 2%to 4%higher.

关 键 词：GABOR特征 几何特征分块对齐 权重化Bp神经网络 表情识别 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程] TP391.41[自动化与计算机技术—控制科学与工程]