Facial Expression Recognition Based On Non-negative Sparse Matrix Factoriation

Expression is an important non-verbal information exchange between people, and is also an important way of transmission human emotions. As an important research topic of emotional computing and intelligent human-machine interaction, facial expression recognition technology has been widespread attention and rapid development in recent years.Usually, Facial expression recognition system is divided into face detection and pretreatment,face feature extraction and expression classification three parts. In this paper the three aspects were studied.The main work is as follows:(1)Summarize the purpose and the significance of facial expression recognition,especially for expression feature extraction and expression classification method of two parts,this paper gives the research status. By reading papers, summarize both at home and abroad since 2006 face expression recognition technology of the latest development.(2)In this paper, the face images detected need preprocessing. We focus on the face image illumination compensation methods. Mainly includes histogram equalization, Gamma correction, homomorphic filter, Isotropic filter and Anisotropic filter five kinds of method.And the Gamma correction is improved. Establishing adaptive Gamma correction function to make every pixel corresponds to a Gamma value, so that improving the light adaptability of Gamma correction and reducing the distortion effect.At the same time, using the block gray integral projection algorithm to position the human eyes. And through the human eye coordinates to correct image, effectively solved the plane offset problem for head rotation.(3)Introduce the basic theory, the objective function and the iterative algorithm of non-negative matrix factorization, local non-negative matrix factorization and non-negative sparse matrix factorization . Through the principle and experimental comparisons, the paper expounds advantages and disadvantages in expression feature extraction of the NMF algorithm, the LNMF algorithm and the SNMF algorithm. And the number of characteristic and sparseness are discussied in experiment..(4)Using class inside block local weighted method to improve the Sparse matrix factorization algorithm, the BWNMF?BWLNMF and BWSNMF algorithm are put forward.Through the block processing, training samples are divided into several subpatterns.According to the different recognition performance for expression recognition to add different weights on subpatterns. Thus give full play to the local characteristics on expression recognition advantage. At the same time, different ways of block are dissussied in the paper.(5)The paper learns the classification principle of the support vector machine. For such multiclass classification problem as facial expression recognition, we introduce three kinds of multiclass classification method, containing "one to all" classification, "one to one"classification and binary tree classification. From the training time and the classification speed two aspects to compare the different performance among the three classifiers. Finally the paper chose the method of support vector machine based on binary tree for expression recognition. In order to make the tree structure is an approximate complete binary tree, the paper is starting from the theory of clustering to train the binary tree.According to above research,the paper set up the facial expression recognition system based on non-negative sparse matrix factorization. Choose JAFFE standard expression library and lab collection expression library to test the facial expression recognition system. The system identification results show that NMF and the improved algorithms are effective in expression recognition.