Research On Face Recognition Based On Feature And Sample Induced Canonical Correlation Analysis

With fast development of face recognition in recent years,the technique of face recognition under normal condition is mature and has been used in many application fields.It can meet actual requirement under the situations of constrained environment and user coordinate.But the current the technique of face recognition still can meet the requirement for effective recognition since the face image exists multi-visual,multi-posture and noise under the complex environment or no user coordinate.For recognition of the face under the situation of multi-visual,multi-posture,it must be to research the method of feature learning and reduction for face image.Canonical correlation analysis(CCA)is an effective method for feature dimension reduction.In the view of this point,the relative improved CCA methods are researched by taking into considerations of requirement of face recognition in complex environment and advantage of CCA.The two improved CCA methods are presented and are used in face recognition in this thesis.The main research contents are described as follows.1)Aiming at the feature to contribute to discrimination of face category,a feature dimension reduction method based on feature induced kernel CCA is presented.In this method,the improved kernel CCA model is built by inducing influence factor to punish the feature.The model is optimized by integrating cosine metric for measuring contribution of the feature and projection vectors,in order to iteratively learn the latent relationship among features and to suppress the influence of the model caused by the different features.The method is used in face recognition.The experimental results show that the proposed method can improve the face recognition accuracy rate under the situations of multi-visual and multi-posture.2)Aiming at the problems of accuracy performance caused by multi-visual and noise,a feature dimension reduction method based on sample induced kernel CCA is presented.The influence factor to punish the sample is induced into traditional CCA,in order to suppress the influence of the model caused by the different samples.The latent relationship between sample and projection vector is iteratively learnt by using cosine metric to measure the projection of sample.Furthermore this method is extended to kernel CCA and the feature dimension reduction method based on sample induced kernel CCA is formed.This method is used to recognize the face.The experimental results show that the proposed method can improve the face recognition accuracy rate under the situations of multi-visual and noise.3)The prototype system of face recognition is designed and implemented by using C++ and MATLAB programming.The system is consisted of three modules,such as face database building,model training and face recognition.The system running shows that the proposed methods are usable for the middle-small scale of face recognition application.