Multimodal Feature Level Fusion Recogniton Based On Canonical Correlation Analysis

With the development of artificial intelligence,single-modal biometrics recognition technology has achieved good recognition results.Due to some problems such as sensor noise and easy to be forged,it is difficult to meet the application requirements in the field of high security.Multimodal biometrics technology combines multiple biometrics for identity recognition.Compared with single-modal biometrics,it is more secure,more practical,and more reliable.Although multimodal biometrics has become a research hotspot in recent years,there are still some issues that need to be further explored.This thesis is the feature layer fusion recognition in multimodal biometric recognition,which fuse the biometrics of two modals and multiple modals respectively.The main research results are summarized as follows:Canonical correlation analysis and its improved algorithm realize the correlation between feature sets by maximizing the between-class scatter matrix between two modes.It does not directly analyze the single-modal feature set,which reduces the recognition effect.Therefore,the correlation between the single-modal feature set and the multimodal feature set is analyzed.First,find the optimal projection direction in which the samples in each modality are concentrated as much as possible,so that the single-mode samples of different classes have greater discrimination.Secondly,two data sets with large discrimination between classes are fused to establish a criterion function,and the optimal projection direction is obtained by using Lagrange multiplication.Therefore,the fused sample set has the largest within-class correlation and the smallest between-class correlation.Experimental results verify the effectiveness of the proposed algorithm.In traditional machine learning fusion algorithms,most of them project multimodal feature sets into a common subspace to achieve the fusion effect.However,the new subspace after fusion has no class information.Therefore,in order to learn a highly correlated,maximum discriminative linear subspace,a feature layer fusion algorithm based on graph-based canonical correlation analysis is proposed.First,based on the L21 regularization algorithm,independent features are selected in each sample space to improve the discrimination of single-modal features and enhance the correlation between multimodal samples.Second,the CCA subspace learning method based on graph embedding establishes a data similarity matrix,which represents the close relationship between the sample points in the original sample space.Therefore,the projected samples maintain the geometric structure of the original sample space,and the between-class samples are as relevant as possible.