基于双支路核化群稀疏学习的微表情识别 |
Micro-expression recognition based on two-branch kernelized groups sparse learning |
|
DOI: |
中文关键词: 微表情识别;稀疏学习;特征融合;特征选择 |
英文关键词:micro-expression recognition; sparse learning; feature fusion; feature selection |
基金项目:国家自然科学基金(72074038,62076122,61971236)和江苏省研究生科研与实践创新计划(KYCX19_0899)资助项目 |
作者 | 单位 | 魏金生 | 南京邮电大学 通信与信息工程学院,江苏 南京 210003 | 卢官明 | 南京邮电大学 通信与信息工程学院,江苏 南京 210003 | 彭 伟 | 斯坦福大学 神经与行为科学学院,美国 加利福尼亚州 94305 | 陈浩侠 | 南京邮电大学 通信与信息工程学院,江苏 南京 210003 | 黄晓华 | 南京工程学院 计算机工程学院,江苏 南京 211167 | 闫静杰 | 南京邮电大学 通信与信息工程学院,江苏 南京 210003 |
|
摘要点击次数: 73 |
全文下载次数: 23 |
中文摘要: |
在微表情识别系统中,常规的特征融合方法会引入冗余或干扰特征,因而会影响识别准确率和效率。针对上述问题,提出一种基于双支路核化群稀疏学习(Two-Branch Kernelized Groups Sparse Learning,TB-KGSL)的特征选择方法,并将其应用于微表情识别系统。首先,提取多个人脸区域的3个正交平面上局部二值模式(Local Binary Patterns from Three Orthogonal Planes,LBP-TOP)和多个方向上的单方向梯度直方图(Histogram of Single Direction Gradient,HSDG)两组不同类型的特征;然后,使用TB-KGSL模型从上述两组特征中分别选择有效区域的LBP-TOP特征和有效方向上的HSDG特征;最后,将选择的LBP-TOP和HSDG特征进行拼接融合,得到紧凑且可鉴别的特征,并使用基于支持向量机(Support Vector Machine,SVM)的分类器进行微表情分类。实验结果验证了TB-KGSL的可行性和有效性,并在CASME II和SMIC数据集上分别达到68.63%和75.95%的识别准确率,比基线方法分别高出5.77个百分点和15.20个百分点。 |
英文摘要: |
In the micro expression recognition system, the conventional feature fusion method will introduce redundant or interfering features and affect the recognition accuracy and efficiency. In this regard, a novel feature selection method based on two-branch kernelized groups sparse learning (TB-KGSL) is proposed and applied to the micro-expression recognition system. First, local binary patterns from three orthogonal planes (LBP-TOP) in multiple facial regions and histogram of single direction gradient (HSDG) in multiple directions are extracted. Second, the TB-KGSL model is used to select the LBP-TOP features in the effective facial regions and the HSDG features in the effective directions from the above two groups of features. Finally, the selected LBP-TOP and HSDG features are concatenated to obtain compact and distinguishable features, and a classifier based on the support vector machine (SVM) is used for micro-expression classification. The experimental results demonstrate the effectiveness of TB-KGSL. The recognition accuracy on CASME II and SMIC datasets is 68.63% and 75.95%, respectively, which is improved by 5.77 and 15.20 percentage points compared with those of the baseline method. |
查看全文 查看/发表评论 下载PDF阅读器 |