Research Of The Ear Recognition Based On The Subspace Analysis And Support Vector Machine

Ear recognition technique is the new method about individual identification, bymeans of the living creature characteristic of human ear. But the researches on earrecognition are all few in domestic and foreign country at present and it is still at thestart.The problems of ear localization, feature extraction and classification arediscussed in this paper. The works that I have done are as follow:1. Ear localization: The speciality of the ear image is that there are not manypose varieties, and the background is not complicated, so we put forward a earlocalization method based on gray-level integration projection and transcendentknowledge of human ear. Firstly, we can get ear's left and right boundaries and itsupper boundary based on this projection theory. Then an ear rectangle outline can beaccurately gotten by means of canny operator to extract ear edge. Finally, through aseries of standardization, the whole ear picture can be given.2. Feature extraction: Because the subspace analysis method can reflect thecharacter of algebra effectively, the ear feature can be extracted based on linearsub-space analysis and on-linear sub-space analysis. This method is proved to beeffectiveness by the further experiments. On the one hand, the methods based on linearsub-space analysis are applied at first.Principal Component Analysis and FisherLinear Discriminate Analysis are discussed in detail and a new algebraic featurebased on using singular value decomposition and projective method is put forward.An approach is also proposed to optimize the recognition performance with least IC(Independent component) features by using genetic algorithm. Both methods based onnon-negative matrix factorization are proposed at last. On the other hand, the kernelmothod is applied into ear feature extraction, and KPCA(Kernel PrincipalComponent Analysis) and KICA (Kernel Independent Component Analysis) arediscussed in detail. The experimental results on the NCU ear databse prove thatcompared with the linear sub-space method, the KPCA feature and KICA featurecontains more useful information in a smaller dimensionality, more robust and more effective.3. Classification machine design: To research ear recognition arithmetic ismainly to design a classification machine. Support Vector Machine (SVM) is usedin this part, and we adopt Muiti-SVM arithmetic,which include one versus restarithmetic,one versus one arithmetic,and direct acyclic graph arithmetic.Experimentsshow the arithmetic is exact and efficient. In addition, we research more aboutparameter of SVM.