融合特征基于深度多核学习的动态表情识别

摘要

针对传统动态表情识别方法由于需要处理多帧表情图片而导致提取的特征维数过高、特征类别较为单一、分类器较难适用异构特征数据等问题,提出在帧数不统一的表情图片序列中,利用慢特征分析自动检测表情序列的峰值帧,继而在峰值帧上分别提取表情的几何特征和Gabor特征后降维,并利用深度多核学习对几何特征和Gabor特征融合后的异构特征信息进行学习并分类,从而提高识别率.在The Extended Cohn-Kanade Dataset (CK+)表情库进行的实验结果表明,识别率可达到94.4%.%In the process of dealing with multi-frame expression images in traditional dynamic expression recognition methods,we usually face some problems which lead to high feature dimension, simple feature category and difficult classifier selection in heterogeneous feature data.In order to improve the recognition rate, we firstly apply the slow feature analysis to detect the peak frame automatically among the expression sequences.Then we extract the geometric features and Gabor features based on the peak frame, and the feature dimension is reduced after Gabor features are extracted.Finally,the Deep Multiple Kernel Learning method is used to learn and classify the heterogeneous features which are fused by geometric features and Gabor features.Experiments show that the recognition rate has reached 94.4%by using the Extended Cohn-Kanade Dataset(CK +).