Font Size: a A A

Reseach Of Finger Vein Authentication Based On Convolutional Neural Network

Posted on:2019-12-14Degree:MasterType:Thesis
Country:ChinaCandidate:H HuFull Text:PDF
GTID:2428330566486949Subject:Engineering
Abstract/Summary:PDF Full Text Request
Finger vein authentication technology as a new biometric authentication technology,has received widespread attention because of its some advantages,such as the high security,convenience and so on.However,there are still some challenging problems in this area.For example,the blurred image leads to a serious decline in recognition accuracy,and the rotation and translation increases the difficulty in intra-class matching.Therefore,a finger vein authentication algorithm based on deep convolutional neural network(CNN)is studied in this paper,aiming to improve the robustness about these problems for finger vein system.In addition,due to the face that the high space and time complexity of CNN make it not suitable for embedded devices,We propose a light-weighting model to compress and accelerate the normal CNN model.the key contributions of this work can be summarized as follows:First,A finger vein authentication algorithm based on deep convolutional neural network is proposed,and it is used to extract the features with strong expressiveness.In recent years,deep learning algorithm has made great achievements in most fields of computer vision.However,the inadequacy of finger vein sample has been limiting the development of deep learning in this field.Therefore,we used a pre-training model to initialize the part of weight of the network,and ultilize some data augmentation strategy to expand dataset.By the way,it can greatly reduce the influence of inadequate sample in finger vein field.Second,a matching strategy similar with the template matching is proposed for matching feature of CNN;and the CNN structure is designed to extract features with spatial information.In this way,the robustness of system for rotation and translation can be greatly improved.As we known,CNN have to learn a lot samples with rotation and translation to obtain the robust feature for rotation and translation.However,the domain of finger vein is unable to provide enough sample so that it is hard to directly obtain this kind of feature.Inspired of the template matching strategy,which deals with the issue of ratation and translation by the matching operation rather than the robust feature,we extract the feature map of CNN as the feature of finger vein and design a corresponding matching strategy based on this feature to improve the robustness to rotation and translation.Third,a idea of light-weighting model is proposed to compress and accelerate the normal CNN model.The normal CNN model is difficult to get well application on embedded devices because of its large number of parameters and high computational complexity.For this reason,we first build a light-weighting network model based on the depthwise separable convolutions,and then train the model based on knowledge distillation algorithm to improve the performance.In this way,the light-weighting model not only has very high compression rate and acceleration ratio,but also basically ensures that the performance is not reduced.Finally,several groups of contrast experiments were conducted on three public databases,and the experimental results demonstrated the effectiveness of our methods.By combining these methods,we can not only effectively enhance the system's robustness to rotation and achieve an excellent performance,but also significantly compress and accelerate the network model to make it more suitable for embedded devices.
Keywords/Search Tags:Biometrics, Finger Vein, Convolutional Neural Network, Light-weighting Network Model
PDF Full Text Request
Related items