The Research Of Feature Extraction Methods And Their Applications

With rapid development of information technology, pattern recognition and artificial intelligence and et al., technologies related to face such as face recognition, pose estimation, expression analysis, 3D facial reconstruction have been widely applied and become research hotspots. However, human facial feature extraction is the prerequisite and basis for these hot issues, it is very critical to study fast and efficient methods for feature extraction. As the two classical statistical models, active shape model (ASM) and active appearance model (AAM) have been widely applied for feature extraction recently, some improved methods are proposed after thorough analysis, these improved methods can improve the accuracy and get more better fitting performance. This research works are listed as follows:(1) In original ASM, PCA is used to extract shape eigenvectors of the training data. Face images are composite consequences of multiple factors such pose, illumination expression and et al. However, the method of PCA addresses only single-factor variations and it is different to separate these factors from original image space, which will lead to a failure fitting of ASM. In order to improve the fitting performance of ASM model, a tensor based ASM model is proposed in this paper. This model decomposes the tensor data containing pose, illumination and expression variations into independent subspace and describes the inner relation among various factors. It can decrease the influences brought by these variations. Furthermore, it is convenient to consider these factors or represent them separately.(2) A new method is used for constructing local profile model of ASM. This new method makes full use of texture information around landmarks, it is very critical to provide a more reliable basis for more precise facial features extraction. In ASM fitting, the best point is obtained by minimizing the Mahanobis distance between template local profile and the local profile of the candidate points in the new image. However, the error is very big when we compute the inverse matrix of a nearly singular covariance matrix using Mahanobis distance. An improved fitting algorithm is used to overcome this flaw. Experimental results show that the two improved methods improve the accuracy of feature extraction.(3) Compared with traditional ASM, traditional AAM includes more rich texture information, because AAM includes global appearance texture information, while local appearance model is constructed by using the neighborhood texture information of landmarks in traditional ASM. However, intensity values used in original AAM can not provide enough information for image texture, which will lead to a larger error or a failure fitting of AAM. In order to overcome these defects and improve the fitting performance of AAM model, an improved texture representation is proposed in this paper. Firstly, translation invariant wavelet transform is performed on face images and then image structure is represented using the measure which is obtained by fusing the low frequency coefficients with edge intensity. Experimental results show that the improved algorithm can increase the accuracy of the AAM fitting and express more information for structures of edge and texture.(4) We wish to fit an AAM by minimizing the error between input image and template image, this is an optimization problem. Generally, the least mean square algorithm is applied to the problem, we introduce some optimization methods used for AAM, and then we compare their performance with different methods by the experimental results.