Feature Vector Optimization In Face Recognition

Biometric identification technology refers to the technology to use human body for biometric authentication. Biological characteristics of the human body are divided into two main categories, One of the categories is physiological characteristics, and the other category is behavioral characteristics . Physiological characteristics includes a series of features of the iris, DNA, palm prints, fingerprints, face, and so on. Voice, handwriting, gait and other characteristics are behavioral characteristics. Biometric identification technology provides a solution for identification recognition, which is now a research hotspot in computer vision and pattern recognition.Face recognition is some kind of technology or system to confirm identity by human faces, and have a wide range of applications in human-computer interaction, biological identity authentication, video detection and video information retrieval. Compared to other identification techniques, face recognition has a advantage of direct, friendly, convenient, non-contact, etc.. At the same time, because of the effect of age, posture, facial expression, illumination and other factors on the face image, face recognition is facing many challenges. In the face recognition, the face image is usually first projected into the feature vector space, which decides the face recognition results. The face image of the same person should be more focus in the good feature vector space, and vice versa.First of all, we give a overall introduction of the background of the face recognition technology development in recent years, and an introduction of the purpose and the contribution of the study in this article. And then on the basis of the relevant literature, we give a overall introduction of the popular face database on the current, and an introduction of face feature extraction and face classification techniques. A concept of classify ability of each vector is put forward from the micro sense. By calculating the classify ability, vectors with greater ones are selected. In order to make vectors with greater classify ability effort more in recognition, they are equilibrated by being given different weights to improve recognition. The experiments on ORL and Yale face database show that the method is efficient.There are two major innovation points in this article:Firstly, PCA(principal component analysis) is one of the most basic the ways for face recognition. PCA algorithm provides a linear transformation matrix from a high-dimensional space to a low dimensional space, and original image is estimated in the sense of the minimum mean square error. It has advantages of high computational efficiency, clear concept , widely application and so on. In the traditional PCA algorithm, we used to select those eigenvectors corresponding to larger eigenvalues to identify, namely, principal component characteristics. But a method of choosing eigenvectors corresponding smaller eigenvalues is also used , that is , the sub-component characteristic. The effect of the two methods are not very good, so a method to choose eigenvectors by their ability to identify and then to weight the eigenvectors according to their ability to identify is prompted to combined with PCA, and finally very good results is achieved.Secondly, we analyze the relationship between the selection method based on identify ability and LDA , and then make further improvements to improve the calculation speed. And further we combine our new strategy and 2DPCA methods. 2DPCA method is based on 2-D images, not 1-D vectors. So this method is more easier, and quickly.The strategy to select eigenvectors based on the recognized ability can better reflect, in the micro sense, each feature vector's effect in the face recognition. According to each vector's ability, they are weighted differently to play a corresponding role in the face recognition. The face recognition system is also widely used, which can be applied in various fields, such as the management of the import and export, identity authentication, social security, intelligent monitor, human-computer interaction, and so on.