Research On Facial Expression Recognition Method Based On Deep Belief Network

Facial expression recognition technology has been widely used in the fields of human-computer interaction,emotion recognition and computer vision,and become a research hot spot in recent years.Some of existing machine learning methods,such as support vector machines,kernel regression,artificial neural network with only one hidden layer and so on,are all shallow structures.However,psychological studies show that the shallow structures are difficult to represent the complex functions in the case of finite samples and computing units.Deep learning algorithm extracts the input data features progressively from the bottom to the top through modeling the human brain hierarchical structure to realize the approximation of complex functions.Thereby,this paper utilizes deep learning method to recognize facial expression.The main work of this paper is as follows:This paper introduces and summarizes the facial expression recognition methods including feature extracting methods,expression classification methods and deep learning.Then,we introduces the basic theories and thoughts of deep learning,and expounds a few classic deep architectures and their learning algorithms,including backpropagation algorithm,restricted Boltzmann machine,deep belief network and convolutional neural network.This paper proposes a novel facial expression recognition approach using firework algorithm and deep belief network.To the best of our knowledge,the traditional deep belief network basically utilizes backpropagation strategy to refine the initial parameters of the deep architecture in the supervised learning stage.Unfortunately,it is easy to fall into a poor local optimum only with backpropagation approach,which is not favorable for deep belief network to find an optimal network parameters.In order to address the problem,this paper fuses firework algorithm(FWA)and conjugate gradient(CG)algorithm to optimize the initial parameters in the supervised learning stage.First,the fusion algorithm utilizes FWA to seek the global optimum of the search space.Then,the CG algorithm is used to optimize the global optimum found by the above FWA to find a better solution.The use of CG operator obviously improves the FWA’s convergence precision and highly accelerates its convergence rate.Our proposed FWA achieves the recognition rates of 91.78%on the JAFFE database and 94.48%on the CK+ database,both of which are higher than the existing methods.This paper proposes a novel facial expression recognition method based on linear discriminant deep belief network.The initial weight matrix between the last hidden layer and the labeled layer of the current deep belief network is usually generated randomly with less discriminant ability,which leads to that the features mapped from the initial weight matrix cannot guarantee to be suitable for classification task.Traditional linear discriminant analysis algorithm has the rank problem,that is,for a train set containing C categories,the rank of the between scatter matrix is bounded from above by C-1,which means that the dimensionality of features extracted by linear discriminant analysis method cannot exceed C-1.To address the above two problems,first,this paper improves the traditional linear discriminant analysis by designing a new between scatter matrix to address the rank problem.Then,we use the improved linear discriminant analysis to initialize the weight matrix between the last hidden layer and the labeled layer of deep belief network to make it suitable for classification tasks.Our proposed linear discriminant deep belief network obtains respectively the recognition rates of 90.41%and 94.48%on the JAFFE database and the CK+ database.