AAM Modeling Based On Gabor-textures

The feature extraction and modeling of the targets, which are the key points of the area of computer vision, have important influences about image understanding and analysis. Active Appearance Models (AAM), advanced by F. T. Cootes et al, is an effective and exact modeling algorithm of feature extraction. This algorithm utilizes Principal Component Analysis (PCA) to train and analyze images, using a statistical way to model both shape and texture of objects. AAM can effectively wipe off the nonlinear coupling between shape model and texture model; reduce the computational complexity of features. So, AAM feature extraction plays important roles in target detection, recognition and tracking. However, as the application expanding, the basic AAM cannot satisfy the needs of accurateness. In this paper, we do a research about the AAM modeling process, trying to improve it form three sides:At first, the texture representation of basic AAM is very simple. In this paper, Gabor filters with good description of scale and orientation are introduced to extract ampler texture features of facial images. These features can help AAM get more accurate facial feature location points. Therefore, the AAM based on Gabor filters can availably improve the feature extraction ability of basic AAM. Secondly, because the Gabor texture features result in the decrease of modeling process, we utilize local Gabor filters bank to represent the image texture in order to reduce the computational. Then, using Genetic Algorithm (GA), we propose to find the best Gabor filters bank to extract the texture features. From these measures, we reduce the complexity of the modeling process of Gabor-based AAM without affecting the accurateness. At last, in order to overcome the disadvantages of vector method, we advance a modeling algorithm based on tensor Gabor feature. Traditional vector method which does not consider the spatial position information currently may destroy the inner data structure of the original data; cannot obtain a compacter model. Therefore, we introduce Tensor Subspace Analysis (TSA) which can reserve the spatial position information of the data to improve calculational efficiency. The experimental results show that the theories in this paper have good performance on AAM modeling.