The Research Of Facial Expression Recognition Based On Multi-LDA

Facial expression is a very important means for interaction between human beings that is complementary to oral conversation. Valuable information can be mined from various facial expressions. In computer vision, facial expression recognition is a process of capturing and classifying the visual expressions into different categories. During this process, the psychological states of human beings can be inferred and therefore facilitates interactions between human beings and machine. The issue of facial expression recognition arises from a very wide range of applications. For example, it has been seen a great demand for facial expression recognition technology from human-computer interaction, machine vision, image processing and pattern recognition. In viewing the great needs from real world as well as the challenges of this problem, this thesis explores effective solutions for this problem. In addition, the related issues have been studied as well.There are three major contributions in the thesis:·An effective feature extraction algorithm from facial expression, namely Skew-Symmetry Local Binary Patterns (SS-LBP) is proposed. This algorithm is built upon two classical algorithms:Local Binary Patterns (LBP) and Center-Symmetric Local Binary Patterns (CS-LBP). Experiments show that SS-LBP performs very well when it is applied on low resolution images, which is very helpful for the image dimension reduction and classification procedures that are undertaken at the afterward stages.·A novel method called Incomplete Weighted Discriminative Filters (IWDF) has been presented. Firstly, a novel pattern of a linear discriminant analysis approach, i.e., Incomplete Weighted Linear Discriminant Analysis (IWLDA), is proposed, which is an extension of the classical Weighted Linear Discriminant Analysis (WLDA) technique. With this discriminative analysis approach, the basic filters are implanted in. Six Incomplete Weighted Discriminative Filters are obtained via IWLDA, which correspond to six different types of facial expressions. The best fit filter is selected from a variety of filters. In the meantime, the best weighting function that is based on Mahalanobis distance is also chosen. Finally the best IWDF is the combination of selected "dot product filter" and the weighting function.·A novel method called Weighted Linear Ridge Regression based on the conventional Linear Ridge Regression method is proposed for image classification. The experiments have been conducted on several popular datasets. The proposed classification approach WLLR achieves much higher precision over the other methods on facial expression recognition.The proposed solutions in the thesis involve the techniques in feature extraction, image filtering, linear discriminative analysis and linear regression.