Study On Multiple Features Fusion Based Face Recognition Algorithm

Automatic face recognition is the process of automatically achieving recognition ofpersonal identity by the computer based on geometric or statistical features derived fromface images. Recently, Face recognition has become a hot research topic in many fields,such as computer vision and pattern recognition, owing to its both significant theoreticvalues and wide range of applications. Extracting the effective representation of facialfeatures from the raw face images is the crucial issue of face recognition. In practicalapplications, due to different imaging factors such as facial expression, pose,illumination conditions, shading, etc, it is not easy to find the features which areefficient, robust and discriminative. It is effective to improve recognition performanceof the representation of facial features to make use of information fusion to analyze andsynthesize the advantages of a variety of facial feature representations. The maincontent of this dissertation are as follows.(1) Technology of face image pre-processing was studied. Firstly, we introduced indetail Adaboost-based face detection algorithm. Secondly, we proposed a novel fastradial symmetry based rapid eye localization, which defines radial symmetry transformto determine the contribution each pixel makes to the symmetry of pixels around it todetect points of high radial symmetry and to accurately locate center of the pupil at alow-computational cost. Thirdly, the method which normalizes the face image wasstudied. Computer simulations validated the effectiveness of the face detectionalgorithm based onAdaboost and the proposed eye location method.(2) Three effectively local facial feature representations, which are respectivelyLBP feature, Gabor feature and Image Gradient Orientation (IGO), were studied. TheLBP operator effectively describes the local texture information of the face images andpossesses the discriminant power. Gabor feature well describes local and globalinformation of face images based on the multi-scale and multi-orientation characteristicsof Gabor filters. Because of its own statistical properties, IGO is robust to the outliers,and especially when combined with subspace learning, IGO achieves high recognitionaccuracy.(3) Considering the deficiency that the facial feature representation ishigh-dimensional vector and contains redundant information, subspace learningalgorithms are utilized to reduce the dimensionality of the facial representation and tocompress noises and redundant information. The experimental results on the differentdatasets validated the effectiveness of the above three facial feature representations，andthen we discussed and analyzed their recognition performance under differentapplication conditions based on the experimental results.(4) Information fusion technology was applied to face recognition. Firstly, we introduced briefly the information fusion strategy applied in multimodal biometricsystems. Secondly, we proposed the subspace reduction of dimensionality basedmulti-feature fusion method. Finally, we applied the above multi-feature fusion methodto fuse LBP feature, Gabor feature and IGO. The experimental results validated themulti-feature fusion based face recognition algorithms can effectively improverecognition accuracy and robustness.