Font Size: a A A

Palm Vein Recognition Based On Deep Learning

Posted on:2020-03-05Degree:MasterType:Thesis
Country:ChinaCandidate:L S YuanFull Text:PDF
GTID:2428330575986690Subject:Biomedical engineering
Abstract/Summary:PDF Full Text Request
With the rapid development of e-commerce,especially online payment,higher requirements are put forward for identity authentication,and traditional identity authentication(based ona password,U shield,IC card)is difficult to meet the requirements of high security,high reliability,and convenience,thus promoting the research and development of biometric identificatioa In the fields of palmprint recognition,fingerprint recognition,vein recognition,iris recognition,face recognition and other biometric recognition,veins(this paper takes palm vein as the research object)have gradually become the focus of biometrics in the field of biometric identification with their unique characteristics of living,uniqueness,rooting and inability to forge.The basic principle is that under the near-infrared light of a palm in a certain wavelength range(wavelength is 720nm?1100nm),the hemoglobin in the subcutaneous vein of the palm has strong absorption characternstics to near-infrared light,and is reflected to the sensor to form a darker pattern,and the surrounding tissue due to weak light absorption,reflected back to the sensor to form a brighter area,In order to form a different brightness of the grayscale image,that is,the image of the palm vein(the following referred to as the "palm vein")can be used for identity authentication after the palm vein image reaches a certain resoluution.Palm vein recognition mainly includes image acquisition,image preprocessing,feature extraction and classification recognition.On the one hand,the traditional palm vein recognition method firstly enhances and denoises the image of the palm vein obtained by near-infrared light,then extracts the vein characteristics,encodes the vein characteristics,generates the eigenvectors that can express the palm vein image,and finally uses various classification algorithms to classify and identify the eigenvectors.Among them,feature extraction is the key step of palm vein recognition.Although the traditional method of palm vein recognition has obtained good palm recognition accuracy,feature extraction relies on the experience of researchers to manually design the feature extraction method which conforms to the palm vein image,and the method also needs to be based on high-quality palm vein image,that is,a lot of operation is needed in the pretreatment stage to highlight the vein information.On the other hand,with the development of convolution neural network in the field of computer vision,people began to use convolution neural network instead of a traditional method to automatically learn the characteristics of palm vein.However,the training of a convolution neural network with strong feature expression ability requires a large number of data samples to adjust the parameters,increase the amount of computation,and a large number of parameters generated during the network training process are prone to over-fitting.In addition,the traditional neural network is mostly single channel,when the network level is small,the feature information of the original palm vein cannot be completely extracted,and the extracted palm vein features are not highly abstracted and the expression ability is not strong,which leads to the recognition accuracy when the palm vein is classified and recognized,when there are too many network layers,it is easy to appear gradient dispersion phenomenon.Based on the above problems,two methods are proposed in this paper:(1)A palm vein recognition method based on two-channel convolutional neural network is proposed.It effectively solves the problem of small sample data,insufficient extraction characteristics of single-channel convolutional neural networiks and excessive gradient of network layers.The method uses multi-channel and different-sized convolution kernels to extract the features of the original palm vein image,obtains rich palm vein feature information,and then performs feature fusion in the fully connected layer to extract deeper classification information.Finally,the softmax classifier is used.Classify and identify.A good result was obtained by using the method in PolyU database,CASIA database and self-built database.The recognition accuracy is 99.90%,90.75%and 95.25%respectively.Compared with the traditional palm vein recognition method,the proposed two-channel convolutional neural network has stronger learning ability and better generalization ability,and can obtain a good result even without pre-processing.In addition,the improvement of neural network in this paper also reflects the good scalability and generalization ability of convolutional neural networks.(2)A new method for palm vein recognition based on transfer learning and random forest is proposed.The palm vein features can be automatically extracted and classified and identified,which avoids the limitation of the artificial selection feature extraction algorithm and effectively reduces the classification error of the palm vein.Firstly,the pre-trained deep neural network model AlexNet is used to extract the palm vein features.The idea of transfer learning is used to effectively avoid a large number of adjustment tasks,improve the classification efficiency of the palm vein,and the feature information extracted by the convolutional neural network model with sufficient depth is more capable of distinguishing.Then the principal component analysis(PCA)is used to reduce the extracted high-dimensional palm vein features to reduce storage space and reduce classification error.Finally,the random forest with good tolerance to noise is classified.The database used for the evaluation experiment was derived from the publicly available PolyU database,CASIA database,and self-built database.The test accuracy was 100%,97.00%,and 99.50%,respectively.Compared with traditional artificial features,the palm vein features extracted by deep convolutional neural networks contain more data information,strong discriminative ability,and are invariant to illumination and rotation.
Keywords/Search Tags:Biometrics recognition, Palm vein recognition, Convolutional neural network, Dual-channel convolutional neural network, Transfer learning, Random forest
PDF Full Text Request
Related items