Research Of Face Recognition Technology Based On Multi Feature Fusion

With the rapid development of network information, information has become a new productivity to promote the development of economy. At the same time, the information security has become a significant factor which has restricted the development of information productivity. The way of rapid and accurate to identify human status has become an urgent problem to be solved. Due to the advantages of non-contact, simple acquisition equipment and multi-platform support, face recognition stands out from many biometric technologies. A real-time face recognition system consists of the modules of face detection, face alignment, image preprocessing, face registration, feature extraction and matching. Face detection is to monitor the video stream to find the location of the face, Face recognition is through the match of detected face images with registered face to the identity human status.In this paper, both the face detection algorithm and the image preprocessing algorithm are introduced in detail. Face recognition is focused on and improved algorithm is proposed. At present, the methods of face recognition based on local feature consist of scale invariant feature transform (SIFT), local binary pattern and Harris corner point. The methods of face recognition based on global feature consist of linear or nonlinear subspace, manifold learning, the neural network and support vector machine. Although these methods have good effects on the condition of the frontal face and simple background environment, but when face pose transformed or angle deflected, the accuracy of recognition will turned fluctuation. It is the main direction to improve the accuracy of recognition under complicated conditions. The main research work is as follows.SIFT is one of the fundamental manner to characterize the local features of human face. In the traditional SIFT algorithm, the coordinates of the feature points are determined according to the local extreme points of Gauss difference space, which introduce the difficulty to match the feature points. To solve this problem, a feature extraction algorithm based on the landmark neighborhood SIFT descriptor (LN-SIFT) is proposed. In this paper, landmark according to human face geometric features firstly can be obtained by the face alignment based on regression of local binary feature. The feature points of the SIFT algorithm are replaced by the landmark which are beneficial to face recognition. Then, the local invariant features of neighborhood around landmark which can effectively improve the identification accuracy are described by SIFT descriptor. Finally, to be robust to the deviation caused by the factors such as the posture change and the angle deflection, the feature of landmark is weighted according to the region of face geometric featuresAs Laplacian Eigenmap (LE) algorithm is difficult to obtain the data of dimension reduction from new samples, locality preserving projection is put forward as the linear expansion of LE. However, this method is an unsupervised method that does not contain the label information, and cannot use the known tag information to improve the identification accuracy. To solve this problem, maximum class repulsion of local discriminant projection (MR-LDP) algorithm is proposed in this paper. The weighted matrix is refined by the local class information, so that the neighbors of the same class samples as close as possible, the neighbor of the different class samples as far away as possible. Meanwhile, to make full use of the global class information, maximum repulsion of different class samples is put forward, so that the deviation of the mean center of different class samples lead to the dispersion of different class samples. Finally, the global feature projection matrix with the manifold structure and the maximum distance between different classes can be obtained.As other higher life, when facing the unknown object, human beings independently extract the most significant features of the object, and then use the integration of characteristics as the proof of identification. Individual feature is difficult to adapt to the change of light, background, gesture and other factors. In order to improve the accuracy of discrimination, a face recognition algorithm based on mixture of the local feature and global feature of human face is proposed in this paper.