Research Of Finger Vein Recognition Based On Deep Learning

Biometric recognition has become an important method of identity authentication.Among numerous biometric identifiers,finger vein attracts widespread attention due to its unique advantage,therefore the finger vein recognition technology developed rapidly these years.However,the existing finger vein recognition algorithms are generally based on handcraft features which are usually sensitive to image quality and finger movements.Aiming at this issue,many improved ideas have been proposed in recent years,which mainly intended to design better image preprocessing methods and more effective handcraft features,but these measures are usually too complicated and the performance of the whole system is still unsatisfactory.In order to learn from original finger vein images,extract more discriminative features,as well as simplify the procedure of image preprocessing,we introduce the method of deep learning,and the main works and contributions of this thesis are summarized as follows:1.Through the use of finger vein image augmentation and the way of building a convolutional neural network based on pretrained weights,we alleviate the issues of lacking finger vein training samples.2.On the basis of pretrained-weights-based CNN,the Siamese structure is constructed for metric learning.Moreover,the Modified Contrastive Loss is proposed for training the network,which makes the performance further improved.By metric learning,the network can extract more discriminative features,but the original Contrastive Loss used in the Siamese network has a strong constraint on the distance of within-class?positive?image pairs,making the network be easily influenced by the extreme samples during training.In fact,this constraint is not necessary for the recognition task,therefore we modify the original Contrastive Loss to loosen the constraint.Besides,this thesis utilizes the multi-stage training with hard examples mining.After the training,the network can extract 256-D features for each finger vein image,and the Euclidean distance-based matching procedure is used in recognition stage.3.Considering that the pretrained-weights-based CNN model is memory-intensive and computationally expensive,which makes them difficult to deploy on embedded devices with limited hardware resources,we further constructed a lightweight CNN model based on the depthwise separable convolution.During the training stage,the knowledge distillation approach is first applied to pretrain the lightweight model,and then the proposed MC Loss is added to optimize the global network.After the training,the weight quantization method is used to further reduce the size of the model.The size of the pretrained-weights-based CNN model proposed in this thesis is 7.95 MB,and the EER?equal error rate?acquired in the dataset MMCBNU₆000,FV-USM,SDUMLAHMT and self-built FV-SCUT is 0.09%,0.15%,0.52% and 1.84% respectively.The size of the lightweight model is only 372 KB,and the corresponding EER is 0.08%,0.11%,0.75% and 1.68% respectively.The experimental results fully proved that the proposed methods have good performance and practical value.