Research On Fusion And Recognition Method Of Finger Biometrics Based On CNN

Multimodal biometric recognition is an identity verification method that combines the features information of multiple modes.Stable and effective fusion method is the key to ensure the application of multimodal biometrics.Fingerprints,finger vein and finger knuckle print have the advantages of compact location,similar biological characteristics and stable ability of identity expression.So,they have gradually become an important research objects in multimodal biometric identification.Deep learning technology represented by Convolutional Neural Network(CNN)has developed rapidly in biometric recognition,which opens up a new research direction for multimodal fusion and recognition.Finger multimodal feature fusion recognition method based on CNN is proposed in this thesis.The specific research content is as follows:1)Recognition model of finger single mode based on CNN is proposed.According to the characteristics of finger features and CNN,we designed an appropriate single mode network which lays favorable foundation for fusion of finger multimodal.2)Feature fusion model of finger multi-modal based on CNN is proposed.First,the convolution feature standardization method is used to unify the features sizes of different modes.Then,three different fusion network models of multimodal features are designed,and the impacts of different fusion layers and fusion methods on the recognition results are analyzed.3)Fusion and recognition method of finger multimodal features based on attention mechanism.Aiming at the problem of large amount of information in features fusion,fusion features of different layers are assigned different attention coefficients.Fusion features with high discriminant are given greater attention coefficient values and relationship among features can be captured.Fusion features with low discriminant are given smaller attention coefficient values to reduce the interference of useless features.With This method the stability and effectiveness of fusion network are improved.