Biometrics technology makes use of human biometrics to achieve effective identity authentication.Compared with traditional identity authentication technology,biometric technology is more secure,easy to carry,and not easily stolen.It has a wide range of applications in areas such as information security.Traditional single-mode biometrics have certain limitations,and multimodal biometrics have higher reliability and anti-counterfeiting capabilities due to their feature diversity and completeness,and can meet the requirements of recognition performance for different applications.At present,most traditional multimodal biometric identification methods have drawbacks:Because the input data of different modalities are quite different,the traditional feature extraction and fusion methods can only learn the correlations among low-level modalities,so it is difficult to obtain satisfying result.Deep learning can make up for this deficiency.The multi-layer neural network structure of deep learning can learn high-level hidden features from the original image,even if the input data has large differences.Deep learning can learn the relationship between abstract modalities so that learns better feature expressions.This article proposes a new multi-modal biological deep learning recognition model based on the stack extreme deep learning machine and the kernel canonical correlation analysis method.The fast extreme learning machine is used as a learning unit to build a deep neural network model to complete the task of multi-modal biometric identification.The main steps of the model are as follows:Firstly,a deep learning model based on extreme learning units,the stack extreme deep learning machine,is constructed to extract high-level feature representations of multiple biometric images.Secondly,using the canonical correlation analysis method based on kernel function to realize the fusion expression of multiple biological features:First,the nonlinear feature representation is mapped to a linear representation,and then the correlation analysis is performed on the two groups of characteristic representations using the canonical correlation analysis.A set of variables which has the largest correlation is the result of feature fusion.Finally,the shallow extreme learning machine classifier is used for training and testing.We test the model on multiple biometric image datasets.Compared with the traditional learning method,the proposed method is more effective. |