Research On Type-2 Fuzzy Face Recogniton

In recent years,the face recognition technology has been developed rapidly,and applied into our life.In practical application,the situations such that the image low resolution,noise,drastic illumination change,and so on occur,because of the influence of various factors during the process of projection imaging.These are all difficult for face recognition and seriously affect the accuracy of face recognition methods.Therefore,how to improve the recognition ability of face recognition methods for the low quality face images is still one of the key issues to be solved.Considering that there is a certain ambiguity in the process of feature extraction and classification,and type-2 fuzzy theory has the advantages of the management of uncertainty and anti-interference,this paper introduces the idea of type-2 fuzzy theory into the research of face recognition,and studies two significant issues in face recognition: the feature extraction and classification identification.The main research results include the following aspects:(1)For complex environment,the change of face images for the same person is often greater than the one between different individual,causing that samples from different individuals have similar characteristics.To address this problem,the type-2 fuzzy linear discriminant analysis(T2FLDA)method is proposed,which introduces the idea of type-2 fuzzy into the LDA method.The proposed method uses the type-2 fuzzy membership degrees to depict the extent to each sample belonging to each class,quantifies the contribution of the different samples to each class in the feature extraction process,and reduces the influence of the outliers.Firstly,the supervised interval type-2 fuzzy C-Means(IT2FCM)is proposed.The clustering centers are initialized using the classification information,and the calculation method of type-2 membership degree is given.According to the classification information,the membership degree of samples is adjusted in order to ensure that the sample is the maximum of its own class.Secondly,T2 FLDA model is built.The fuzzy within-class scatter matrix and fuzzy between-class scatter matrix are defined,respectively.Based on Fisher discriminant criterion,the best fuzzy projection matrix can be achieved.Finally,we analyse the significance of the parameters and the impact on the recognition rate,and give the appropriate choice of parameters.Experimental results show that the obtained embedding subspace has discriminant and robustness.(2)In view of the relatively weak theoretical foundation of fuzzy classification methods,especially the low generalization ability problem,the type-2 Takagi-Sugeno fuzzy classification system based on support vector machine(T2T-SFCSVM)is proposed in this paper.Firstly,the type-2 T-S fuzzy classification system is built,and a decision function of T2T-SFCSVM is designed to solve the problem of binary classification.Secondly,by using the fuzzy iterative self-organizing data analysis technique(FISODATA),we achieve the number of fuzzy rules and the centers of type-2 fuzzy Gaussian membership functions.The uncertain widths of type-2 fuzzy Gaussian membership functions are learned using particle swarm optimization(PSO),and the calculation method of the value of optimal parameters is given.This is done to reduce the influence of artificial factors.Based on the consequent parameters learned through SVM,the consequent parameters calculation formulas are derived.A supervised learning algorithm is proposed for studying the parameters of the plant inference engine,and the out error function of T2T-SFCSVM is defined.Finally,the T2T-SFCSVM is generalized to multiple classification method using“one-to-one”strategy,and then is applied to the identification.The experimental results show that the performance of the proposed method is superior to that of the previous mainstream classifiers.(3)For the case that twin SVM(TWSVM)is sensitive to noise and outlier,the type-2 T-S fuzzy classification system based on TWSVM(T2T-SFCTWSVM)is presented.Firstly,we build the T2T-SFCTWSVM,and design two decision functions of T2T-SFCTWSVM to solve the problem of binary classification.Secondly,the consequent parameters calculation formulas are derived based on TWSVM.Finally,we generalize T2T-SFCTWSVM to the multiple classification method using“one-to-one”strategy,and then apply it to the identification.The experimental results show that in the case of large samples,T2T-SFCTWSVM method requires less parameters learning time and get higher recognition rate;while in the case of small samples,T2T-SFCSVM method requires less parameters learning time and get higher recognition rate.In conclusion,these two kinds of fuzzy classification methods are superior to SVM and TWSVM in recognition rate and robustness.(4)Face images are not only less sample size,but also easy to be affected by various factors,which makes it more challenging to obtain over complete dictionary,dictionary learning and solving the sparse coefficient in sparse representation based classification algorithm.In order to address this problem,this paper proposes the type-2 fuzzy fisher discrimination dictionary learning(T2FFDDL)method.Firstly,we present the T2 FFDDL model.By combing the classification information to construct a structured dictionary,the type-2 fuzzy fisher discrimination criterion is introduced into the coding coefficients to minimize small fuzzy within-class scatter but maximize fuzzy between-class scatter,which makes it discriminative and robust.Secondly,it can be proved that the type-2 fuzzy discriminative constraints function is strictly convex and differentiable on the domain,which provides a theoretical basis for converging to global optimal based on T2 FFDDL.By iteratively optimize the T2 FFDDL model,the desired discriminative dictionary and type-2 fuzzy discriminative coefficients can be obtained.Finally,the classification scheme is presented to make the recognition results more stable.The experimental results show that the proposed method can reduce the influence of interference of atoms in sub-dictionary,obtain complementary information from similarity atom in different sub-dictionary,and can achieve better recognition rate and robustness than other previous mainstream sparse algorithm with small dictionary.