基于多特征决策级融合的表情识别方法  

Facial Expression Recognition Method Based on Multi-feature Decision-level Fusion

作　　者：黄忠[1,2] 胡敏[2] 刘娟[1] 

机构地区：[1]安庆师范学院物理与电气工程学院,安徽安庆246011 [2]合肥工业大学计算机与信息学院,合肥230009

出　　处：《计算机工程》2015年第10期171-176,共6页Computer Engineering

基　　金：国家自然科学基金资助项目(61300119);国家自然科学基金资助重点项目(61432004)

摘　　要：为实现多源特征的优势互补并融合多分类器的决策结果,提出一种改进的多特征表情识别方法。利用链码编码刻画表情形状特征并构建形变特征描述面部几何变化,构造Gabor特征融合图以表征表情局部纹理细节。采用支持向量机分类器分别获取3类特征的类别后验概率并在决策级实现多分类器的融合。在有监督学习下提出一种基于粒子群算法的权重寻优策略求解最优融合权重。Cohn-Kanade表情库上的实验结果表明,该方法在平均识别率和鲁棒性方面均优于单分类器识别方法,与现有的多分类器融合方法相比,权重寻优策略在识别率和可靠性方面更优。In order to perform advantages of complementary multisource features and fuse decision results of multiple classifiers,a multi-feature facial expression recognition method based on decision-level fusion is proposed.Shape Feature（SF）of expression is attained by chain code and deformation feature is built to depict facial geometric changes.Meanwhile,Gabor feature fusion diagram is applied to describe local texture details of facial expression.The posterior probability of three kinds of features,which is obtained by Support Vector Machine（SVM）classifier respectively,is constructed for multiple classifiers fusion in decision-level.In order to solve the optimal fusion weights,a weight optimization strategy based on Particle Swarm Optimization（PSO）under the condition of supervised learning is put forward.Experimental results on Cohn-Kanade database show that the proposed method has better performance for average recognition rate and robustness than single classifier recognition method.Compared with existed multiple classifiers fusion methods,the weight optimization strategy has advantages in terms of recognition rate and reliability.

关 键 词：决策级融合 主动形状模型 链码 形状特征 形变特征 GABOR纹理特征 粒子群寻优 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]