Research On Finger Knuckle Print Recognition Based On Deep Learning

Finger knuckle print recognition is a new kind of biometric technology in today's society,which is determined by individual genes and has individual differences,so it can be used in identity recognition technology.Compared with common fingerprint recognition,the knuckle print feature is not as easy to wear,sweat and moist as the fingerprint,and does not leave contact marks everywhere in the daily life like the fingerprint,which improves the safety of the biological feature.Compared with other biological features,the knuckle print features stable performance,easy collection,and easy integration with other hand-shaped features,and compared to human features such as face and sound,it involves less privacy and is more acceptable to people.Therefore,knuckle print biometrics recognition technology shows a good development prospect.In this paper,the open-source HKPU-FKP finger knuckle print data set is used.Aiming at the problems of traditional machine learning algorithm,such as complex feature extraction,single feature and low recognition accuracy,this paper uses the convolutional neural network model that can extract image features by itself.Research on the task of finger knuckle print recognition and classification,the main contents are as follows:(1)pre-processing of knuckle print images,including ROI extraction of the region of interest,data set expansion,histogram equalization and other operations,reduces the size of input data,shortens the training time,and is conducive to texture feature extraction and inhibition of network overfitting.(2)A 9-layer convolution neural network model is designed and constructed.After compiling and training the network model,the serious over fitting phenomenon of the network is analyzed.Aiming at over fitting,a scheme of optimizing the network structure is proposed.On the basis of the original network model,the network structure is optimized by adding dropout layer,L2 regularization layer,batch normalization layer and other operations,and the experiment is carried out.The results show that the over fitting phenomenon is obviously improved,the recognition accuracy of the network model on the test set is improved,and the generalization ability of the network is enhanced.(3)In view of the characteristics of the finger knuckle print belonging to a small data set,instead of using the expanded data set,we adopt the transfer learning scheme suitable for the small data set,and use the Vgg-16 network pre-trained on the Imagenet data set to carry out the transfer learning on the finger knuckle print data set,and carry out the joint training on the partial layer thawing of the pre-trained network and the classifier.After fine-tuning optimization,the pre-trained network model is more suitable for this recognition and classification task,which improves the portability of the pre-trained network,so that the recognition accuracy of the network model on the test set can be further increased,and enhances the generalization ability of the pre-trained network on this task.