Research On Facial Expression Feature Extraction Algorithm Based On Subspace Analysis

Facial expression can provide the sense clue of human emotion, thus facial expression recognition has been paid extensive attention of many researcher as a key function of human computer interaction. Because of non-rigid feature of facial expression, it is difficult for the traditional facial expression recognition methods to achieve satisfactory performance. Subspace analysis method is a research focus of facial expression recognition, which utilizes statistical analysis technique project samples into the optimal subspace, for completing dimension reduction and feature extraction. This dissertation focus on some issues on feature extraction of facial expression, aiming at extracting effectively represent characterization of facial expression for classification and recognition. The main contributions of this dissertation are tabled as follows:Firstly, aiming at the problem of feature extraction methods based on different criterion cannot extract the effective discriminant information, generalized multiple maximum scatter different(GMMSD) criterion and the corresponding feature extraction algorithm are proposed. GMMSD employs the difference criterion instead of entropy criterion, which can avoid the small sample size problem, and extract more effective discriminant feature for facial expression recognition and face recognition. At the same time, it reduces the computation complexity of feature extraction. Compared with the traditional feature extraction methods, GMMSD contains the following advantages of:(1) GMMSD avoids to the small sample problem, and dispense with preprocessing;(2)GMMSD extracts the discriminant information by QR decomposition, which preserves the distribution characteristics of the original samples;(3) According to different transformation matrices, GMMSD can evolve to different feature extraction algorithms. In other words, GMMSD has the feature of generalization. Experimental results show that GMMSD can provide favorable recognition performance with high computational efficiency.Second, aiming at the possible sample(s) add from training set, we put forward an incremental version of GMMSD, which called IGMMSD+. The incremental version could be divided into two cases: inserts new samples from a new class and inserts new samples from an existing class, and two incremental updating algorithms are given respectively. IGMMSD+ can avoid the re-computation of the whole GMMSD+, when additional training samples are presented. The key features of IGMMSD+ are:(1) its ability to accurately and efficiently update the discriminant vectors with new training samples(more than one sample) incrementally;(2) By employing a projection from both range of centroid matrix and null space of within-class scatter matrix, the method divides the centroid vector of each class into intrinsic common component(ICC) and discriminant difference component(DDC), and automatically discards the ICC with little discriminative information, while keeping the DDC with true discriminative power;Third, to better reveal the existed potential structure of facial expression, multiple-manifolds discriminant analysis(MMDA) is proposed. Traditional manifold learning methods assume that all training samples inside on one common manifold. However, this assumption cannot always hold in many real world scenarios. In other words, we cannot be sure that all the samples of training set on the same manifold. Unlike these manifold learning methods, considers that not all local patches have meaningful positive influence on the formation of facial expression, MMDA models one manifold for each expression using only salient regionsto extract expression-specific features. To sum up, MMDA has the following two advantages:(1) MMDA only utilizes salient regions to build the training and test set. This way not only avoids the interference of no-expression regions, and reduces the computation complexity of the algorithm simultaneously;(2) Compared to the single manifold expression recognition algorithms, MMDA not only avoids the over-fitting problem because of the limitation on the number of training sample, but also improves the recognition performance of facial expression.Finally, most methods converts the images into one-dimensional vector, it may casue the problem of high dimension, computation complexity and the small sample. To overcome these problems, two dimensional multiple-manifolds discriminant analysis(2DMMDA) is proposed. 2DMMDA employs two dimensional data to extract the feature of facial expression, and avoids the problem of small sample size problem. It utilizes 2DMMDA criterion to model the manifold structure of each expression by maximizing the manifold margins of different expressions and minimizing the manifold margins of the same expression, simultaneously keeping the diversity information of samples from the same expression. Two effective implementations are also performed to compute the 2DMMDA criterion. Extensive experiments are carried out to compare the difference between one dimensional and two dimensional facial recognition methods based on independent-subject, and show that the proposed method significantly improves the recognition performance of human emotion.