Research Of Face Recognition Algorithm Based On Feature Fusion

Face recognition is a biometric technology which is great developable potential, researching on the face recognition technology, and it has great theoretical and practical values. In recent years, face recognition technology has achieved unprecedented progress. However, most of the existing face recognition algorithm can hardly achieve the required accuracy, that’s the reason why it has not been widely used in real life. Especially under the non-ideal conditions (such as variations of illumination, photographing azimuth or other disturbance the presence of image acquisition) and faces to the massive database, the recognition accuracy will decline sharply. Under the background of the National Science and Technology Support Program, this thesis is based on feature fusion and makes some studies as follows:First, it takes further analysis of the current face recognition technology research status, existing difficulties and the future direction. At the same time, the preprocessing of face image is researched. The aim of face image preprocessing is to standardize the face image which is captured by image collecting devices to normalized image, it mainly includes three steps:face detection and eyes location, geometry normalized, grey value normalization.Second, some common methods of face recognition is researched. By fully consider the pros and cons of these feature extraction algorithm:Gabor wavelet transform, Local Binary Pattern (LBP), the local phase quantization (LPQ), principal component analysis (PCA) and linear discriminant analysis (LDA), the features fusion method is proposed. First, when the local feature extraction algorithm calculates the histogram based on block, it uses a variety of block methods to an algorithm, making up the losing information between the block and the block. Second, we propose a global feature extraction algorithm to reduce the dimension to the characterize obtained by the local feature extraction algorithm, which removes the redundant features information, enhancing the capability of classification. Third, the feature extraction algorithm of LBP, LPQ and Gabor wavelet are fused, making use of the advantages of the three algorithms and achieving better recognition. The thesis put forward the overall fusion method based on the three different fusion strategy. In order to obtain the overall contribution of the various components of the fusion method, the thesis makes use of Genetic Algorithms and gets the best weights.Third, the proposed algorithm based on feature fusion is tested in the massive uncontrolled database FRGC and receives the preferred recognition rate of 96.45%, meeting the requirements of the National Science and Technology Support Program for the the National Science and Technology Support Program.