Research On The Technology Of Face Recognition With Single Sample Based On Generated Virtual Image And Fusion HMM

Face recognition technology has been attached great importance to the researchers for its scientific significance and practical value in the past few years,and become the hotspot of current pattern recognition and artificial intelligence.Face recognition normally be regarded as have three processes that are face detection,features extraction and pattern classification.Face recognition often meet these problems,the dimension of sample too high,the classes of pattern too many and each person could only provide a single training sample.In this paper,these problems have been studied.By systematically analyzing of relevant algorithms,we present some novel algorithms for face recognition with single sample based on Fusion-HMM and generated virtual image,from two aspects of features extraction,virtual image generation and face recognition.In addition,a prototype system of face recognition is designed and implemented.The highlights and main contributions of the dissertation include:(1) A novel method based on Bilateral Two-Dimensional Linear Discriminant Analysis and Symmetry average of Local Singular Value Decomposition for face recognition is presented. Firstly,Bilateral 2DLDA was used to extract feature on the whole face image;secondly,symmetry average of Local-SVD is used to extract the local facial features;finally,by fusing the extracted features from both two methods above,the nearest neighbor classifier based on the weighted-Euclidean distance is employed to accomplish the task of classification.This method can extract the optimal discriminant features,and make the best of both the whole and local features of the image,as well as get over the influences effects of illumination,expression and gesture to some extent,and be also of great help to the feature extraction problem in single sample case.(2) An algorithm of three-layers Virtual Image Generation is proposed.This algorithm analyzes the advantages and disadvantages of existing algorithms of Virtual Image Generation, combines with geometric transform algorithm,algebra transformation algorithm and the spatial distribution algorithm.Firstly,It uses SVD-perturbation to highlight the facial features;secondly, applies the method of geometric transformation to enhance the changes of attitude and scale, increases the number of sample;finally,uses the method based on spatial distribution to improve the distribution of samples,so that the distribution of virtual samples is more close to the real-world distribution.(3) An algorithm of Fusion-HMM is proposed.For traditional singular value feature contains less information and LDA feature is sensitive to the geometric changes,the B-2DLDA feature is overall and the SL-SVD feature is local,they complement each other well.Based on the information fusion theory,a Fusion-HMM is provided,which is made up of SL-SVD-based HMM and B-2DLDA-based HMM so that B-2DLDA and SL-SVD can complement each other.Both of them are fused by different weight to improve the final performance.The experimental results have shown that the approach is prior too traditional HMM,as it is easier and better performance.(4) Based on the idea of object-oriented,we design and development a prototype system of face recognition with single sample,which is divided into four modules that is image preprocess, virtual image generation,facial feature extraction and face recognition.And makes the system recognize people according to face image when the training samples with little or only a single sample,System has managed to maintain a higher correct recognition rate.