Research On Image Recognition Algorithm Based On Fuzzy Feature Extraction

In image recognition,the same pattern can be described by different features from different angles.Generally,these features can reflect the different characteristics or views.Canonical correlation analysis(CCA)is a classical joint feature extraction method for multi-representation data;it aims to find two groups of projection directions,which can make two groups of canonical projections maximally correlated.Essentially,CCA is an unsupervised subspace learning algorithm,which does not utilize the class label information.Meanwhile,the theory of fuzzy mathematics believes the boundary between sample and classes is not very clear,which means the same sample can belong to more than one class in different degrees.Combined with fuzzy mathematics,a series of fuzzy feature extraction algorithms are proposed on the basis of CCA,and applied to image recognition.In general,the main content and innovative results are as follows:(1)A novel fuzzy CCA algorithm(CCA_f)is proposed.The proposed algorithm describes the fuzzy membership between sample and class from different vies,which makes the information of this fuzzy relationship more completed.On this basis,build the fuzzy CCA model,and the optimal solution is provided.The experimental results on COIL-100 object database and AR face database show the fuzzy feature extracted by our proposed method is more discriminative in image recognition.(2)The fuzzy fractional CCA algorithm(FFCCA)is proposed.This algorithm re-estimates the fuzzy covariance matrix using fractional-order idea,which makes the matrix closer to its real value.The experimental results indicate that FFCCA can achieve two percent higher than CCA_f when the number of training sample is not "enough",which makes the reestablish of fuzzy fractional scatter matrix meaningful.(3)The fuzzy locality preserving CCA(FLPCCA)is proposed.It learns the high-dimensional multi-representation data with locality preserving idea,which transforms the global nonlinear problem into a collection of local linear problems.The feature extracted in this way can expose the essence of nonlinear data better.The experimental results prove that our proposed method has better recognition results.