可鉴别的多特征联合稀疏表示人脸表情识别方法  

Face Expression Recognition Method Based on Discriminative Multi-features Joint Sparse Representation

作　　者：丁泽超 毛启容[1] 詹永照[1] 王敏超[2] 

机构地区：[1]江苏大学计算机科学与通信工程学院,江苏镇江212013 [2]江苏国光信息产业股份有限公司,江苏常州213015

出　　处：《小型微型计算机系统》2016年第12期2775-2779,共5页Journal of Chinese Computer Systems

基　　金：国家自然科学基金项目(61272211)的资助

摘　　要：针对目前多种特征联合的方式多为线性组合方式的问题,提出一种融合图像纹理特征和全局位置特征的可鉴别的多特征联合稀疏表示人脸表情识别方法.该方法首先获取人脸图像的纹理特征和全局位置信息,构建训练字典,通过引入鉴别损失函数,优化稀疏表示的字典.然后在惩罚函数中引入基于类级联合稀疏正则项,对局部纹理特征和全局位置特征进行联合稀疏表示,最后将稀疏表示获得的稀疏系数矩阵送入支持向量机进行训练和表情识别.在Curtin Faces和BU_3DFE人脸表情库的实验结果表明,该方法在降低联合特征维度的同时,能够挖掘多种特征之间的关联性,并使得学习到的稀疏系数更具鉴别性,与新近的多特征融合人脸表情识别方法相比,对BU_3DFE库的7种情感上面平均识别率提高了2.5%-5%.To solve the problem that the combination of multi-features are linear. We propose a facial expression recognition method with discriminative multitask joint sparse representation based on local image texture feature and global coordinate information in this paper. First, Texture features of the image and global coordinate information are achieved to construct training dictionary, discriminant loss function are introduced to optimize the dictionary. Second, introduce a joint sparse Regularization term in objective function to joint sparse representation. At last, the sparse coefficient matrix after sparse coding can be used for training and expression classifica- tion in SVM. The proposed method is extensively evaluated on Curtin Faces Database and BU_3DFE Database. The experimental re- suits show that this method can reduce feature dimension and mining the correlation between multi-features, and enhances the power of discrimination of Sparse coefficient matrix. In compared with the state-of-the-art methods in expression recognition, the proposed meth- od has about 2.5% -5% improvement on the BU_3DFE Database.

关 键 词：多特征 表情识别 联合稀疏表示 特征融合 

分 类 号：P391[天文地球—地球物理学]