Research On Face Image Recognition Based On LDA And Sparse Representation

Face recognition is an important branch of artificial intelligence,and it is of great significance in civil and commercial fields,such as human-computer interaction and monitoring systems.In recent years,with the rapid advancement of machine vision and machine learning,the design of feature extraction and classifiers has become increasingly complex and diverse.However,factors affecting face recognition such as small samples have been eliminated difficultly,and these have made the requirements for feature extraction and classifiers more and more rigid.Based on the research of machine learning algorithms,this thesis is carried out researching on key techniques of face recognition,classifier design and feature extraction are involved.And under small sample environments,classification algorithms and feature extraction are implemented.The research results as follows:First of all,Research and implementation of principal component analysis and feature extraction method based on linear discriminant analysis are done.Feature extraction by improved linear recognition analysis algorithm can solve small sample problems well.Secondly,Aiming at the shortcomings such as variance and deviation of eigenvalues in existing linear discriminant analysis algorithms,a weighted smooth deterministic linear discriminant analysis algorithm(WSDLDA)is obtained by weighting and smoothing the inter-class and intra-class dispersion matrix models and making the normalization parameters deterministic.The improved algorithm can not only solve the over-fitting problem caused by small samples,but also solve the singularity problem of scattering matrix.Three public databases were used to verify the superiority of the proposed algorithm.The performance of the proposed algorithm under different training sample numbers and dimensions are analyzed,and compared with various related methods.The results show that the proposed method can improve the classification accuracy.Thirdly,several commonly used classifiers are deeply studied.Combined with respective advantages and disadvantages,the classification recognition method based on sparse representation is studied.It does not require a large number of training samples,and is easier to train,and can better prevent the over-fitting problem of small sample data.Aiming at the shortcomings such as non-supervised classification of sparse classification,a weighted classification sparse and local sensitive representation algorithm(WCSLR)is proposed,which makes full use of related classes and related samples of test samples.The algorithm not only promotes sparse representation on categories,but also enhances the local sensitive representations in selected training classes.Finally,the same sample data as the previous feature extraction are used.The classification performance of the classifier under different training sample numbers and dimensions is simulated and analyzed.Comparing the classification results of multiple classifiers,the proposed method,WCSLR,has more excellent classification performance is verified.Parameter analysis of the proposed method is also performed.Based on the above research,feature extraction and face image classification algorithm is implemented,and the superiority of the algorithm is verified on the public face database.