Palmprint Recognition Based On Convolutional Neural Network And Transfer Learning

Nowadays,there are more and more scenarios that require identity authentication.Safe and efficient biometric identification technology has been widely used in the fields of intelligent security,identity verification,and judicial authentication.Palmprint recognition technology has become an important technology in the field of biometrics since it has the advantages of stable and rich feature information,simple acquisition equipment,and high user acceptance.Traditional palmprint recognition technology needs to design a feature description algorithm based on texture information to extract palmprint features and then use the classifier to identify.Convolutional Neural Network(CNN)combines feature extraction and classification recognition into an end-to-end identification method,which can adaptively learn and understand the information in an image.A large number of training samples are required by using CNN for identification,but the number of palmprint samples collected in the same category is generally limited,which is not enough to drive effective learning of CNN.Considering that the palmprint image is susceptible to noise pollution during sample collection and the palmprint image is easy to shift slightly when extracting the region of interest,this paper proposes a palmprint recognition method based on multi-wavelet for data augmentation and CNN.When the palmprint library changes or the number of categories is large,the cost of retraining CNN is very high.To reduce the training cost,pre-trained Deep Convolutional Neural Network(DCNN)is transferd to palmprint recognition.And then pruning compression and quantization compression is used to remove the redundancy of the deep network.In order to further improve recognition rate and robustness of the transferd DCNN,the data enhancement method proposed in this paper is applied to the preprocessing of transfer learning,and a palmprint recognition method based on data augmentation and transfer learning is proposed.The main works are as follows:(1)A palmprint recognition method based on multi-wavelet and CNN is proposed.Firstly,use the data augmentation method based on multi-wavelet transform to increase the number of image.Every palmprint image is clipped into five sub-images in five directions.Three components corresponding to the prefilters in the CL low frequency component of clipping sub-images and reduced sub-images constitute the augmented image library.Then,CNN is trained by using augmented palmprint images.The network structure(number of convolution kernels,size of convolution kernel)and the hyper parameters(learning rate,batch size,and epoch)are established through experiments.The results of experiments show that the constructed CNN by this method has better recognition effect and robustness only by fine-tuning some parameters for different numbers of training samples,training samples and palmprint libraries.(2)An optimized palmprint recognition method based on transfered DCNN is proposed.Firstly,compare the generalization ability of different pre-trained DCNN on palmprint images,extract the deep convolution features of palmprint images in pre-training models AlexNet and VGG19,perform PCA(Principal Component Analysis)to reduce the dimension of features,and then use the nearest neighbor classifier to classify.According to the classification result,select AlexNet with better performance and shallow network layer to transfer.Then,Fine-tune the pre-trained AlexNet with different learning rates and get a fast convergent network for palmprint recognition.Finally,to further optimize the fine-tuned network,use pruning compression to remove small weight and enhance large weight,which makes the network sparsed,and then use quantization compression to reduce weights representation with fewer bits,which makes the network lighter.PolyU library and CASIA library are used in experiments.The experimental results show that the recognition ability of fine-tuned DCNN is better than traditional palmprint recognition algorithms,and the recognition rate of compressed DCNN after pruning and quantization is better than ones only fine-tuned DCNN.(3)Combine the advantages of two methods in(1)and(2),a palmprint recognition method based on data augmentation and transfer learning is proposed.The palmprint image is augmented according to the method in(1),and the transferd AlexNet is retrained on the augmented image library.The PolyU library and CASIA library are used the proposed method.The experiments show that that the transferd DCNN with augmented image has higher recognition rate and stronger robustness than the transferd DCNN with original image.