The Embedded Finger Vein Recognition System Based On Convolutional Neural Networks

With the rapid development of economic and technological,more and more information products and equipment appear in people's daily life.Therefore,information security is particularly important.The traditional personal authentication methods have been unable to meet the demand of information security in modern society.Now more and more researchers are focusing on biometrics recognition technology.And as a member of biometric features,compared with other biometric features,finger veins exhibit some excellent advantages in application.For instance,small image acquisition device,low cost,non-contact acquisition,multi-finger can be used.However,at present,the finger vein recognition still has the problems such as the low quality of the acquired image,the spoofing attacks,and the low robustness of the finger rotation along the axis.These problems have a great impact on t he system's performance and user experience.Aiming at these problems,we proposed a new recognition system in this paper: the embedded finger vein recognition system based on convolutional neural networks,which has the function of spoofing detection and image quality assessment.By compared with the existed recognition systems,our contributions are summarized as follows.First,we proposed a new spoofing detection method,and the method was embedded into the recognition system.Aiming at the printed spoofing-attack finger vein image,firstly,we used the Butterworth high pass filter to extract the high frequency image information,and then the LBP method was used to extract the texture features of the high frequency image.Finally,we used the SVM classifier to classify the image.This method can effectively solve the problem of spoofing attack.Second,we proposed a novel finger vein image quality assessment method.There has been a predominant issue for the finger vein identification system that the quality of images is poor.We also proposed a scheme of multistage intensity image acquisition,but the scheme cannot completely solve the problem of poor image quality.Aiming at the characteristics of the finger vein image,we evaluated the quality of images by counting the real vein points,and then filtered the low-quality images.Third,we proposed an improved method based on t raditional identification method.Aiming at the problem of finger rotation along the axis,we used template matching method to correct the region of interest of the finger vein image,and then extracted the texture and orientation features,finally used weighted fusion method to fuse the features matching scores.This method can solve the problem of finger rotation to a certain extent.Fourth,we proposed a feature extraction method based on convolution neural network.In this paper,a simplified network model is designed.Experimental results showed the proposed method outperformed the traditional feature extraction method and achieved the strong robustness of the finger rotation along the axis.In addition,we constructed two finger vein image databases for the need of research.One is with rotation characteristics,and the other is the spoofing attack database.Several groups of contrast experiments had been conducted on the self-built database and the public database,which finally verified the effectiveness of our proposed recognition scheme.Finally,the proposed recognition system gained an Equal Error Rate(EER)of 5.42% in the case of the finger rotation along the axis,which extracted feature based on convolution neural network and had the function of spoofing detection and image quality assessment,and the value of EER was 17.619 % lower than the fusion method based on texture and orientation feature extraction.The result demonstrated the effectiveness of our system.