基于LBP/VAR与DBN模型的人脸表情识别

Facial expression recognition based on LBP/VAR and DBN model

针对现有表情识别研究中均采用有监督模型实现特征提取,提出一种新的基于DBN(deep belief net)模型无监督的表情特征提取与识别方法。首先通过对人脸表情图片提取对光照与旋转具有鲁棒性的LBP/VAR初次特征,再通过DBN网络对初次特征实现人脸表情的二次特征提取与分类学习。对DBN参数采用动态搜索的方法,即在一个大范围内搜索确定RBM Mini-batch、BP Mini-batch与RBM隐层数量的最优值,再确定DBN深度与迭代次数最佳值。在CK+数据库上与传统KNN、SVM有监督分类模型进行的对比实验表明,提出的方法在识别率上分别提高了19.34%和14.22%。

In contrast to the supervised feature extraction method adopted in facial expression recognition, this paper proposed a new method based on the deep belief net model of deep learning architecture. Firstly, it extracted the LBP/VAR feature to forn~ the first feature because the LBP/VAR feature was robust to the light and rotation. Then used DBN model to extract the second feature and to implement the classification of facial expression. For the DBN＇ s parameters, this paper set a wide dy- namic range to search the proper value of RBM Mini-batch, BP Mini-batch and number of hidden unit, then in search of the best value of DBN＇ s depth and the number of epochs. It took the experiment on CK ＋ database and the new method had an ex- cellent performance. The result shows that the recognition rate of this new method increases 19.34% and 14.22% contrast to KNN and SVM.