Research And Implementation Of Data Augmentation And Multi-Task Learning Algorithms In Finger Vein Recognition

Compared with traditional authentication methods,biometric technology has the advantages of convenience and high security.In recent years,biometric technology,which has been developing faster and faster,has gradually replaced the traditional key,password,magnetic card and other security authentication methods.Finger vein recognition technology is an important branch of biometrics.Compared with biometrics such as face recognition and fingerprint recognition,finger vein recognition is superior in terms of stability and accuracy,and has higher application value.However,the field is still in the initial stage of development.The effects of ROI(Region of Interest)extraction and feature extraction of vein images are not satisfactory.Recently,deep learning has gained rapid development due to the explosion of big data on the Internet,and it has surpassed traditional algorithms in the field of images.This paper combines deep learning to explore the problems that still exist in the field of finger vein recognition.And put forward corresponding solutions.The work of this article is as follows:(1)Aiming at the problem of small amount of data in the current public data set,an image quantity enhancement algorithm based on deep learning is designed and implemented.The algorithm uses a conditional generation model to treat finger contours as conditional inputs and finger categories as semi-noise inputs.It trains a data generator that can generate different types of contour finger vein maps.In this paper,we use this generator to generate a vein image dataset containing 2400 categories and name it FV-BUPT.Subsequent experiments evaluated the quality of the image by using a non-reference image quality evaluation algorithm,and verified the value of the data set as training data and pre-training data on the vein recognition model.(2)Aiming at the problems of poor robustness of traditional ROI methods,complicated data feature processing flow,and poor feature extraction quality,a multi-task learning-based finger vein feature extraction algorithm was proposed.This algorithm combines ROI extraction task and feature extraction task to form a multi-task neural network model.Train and run with End-to-End.Experiments show that the model proposed in this paper is superior to traditional algorithms in terms of operating efficiency and quality of extracted vein features.(3)Design and implementation of finger vein authentication demonstration system.This system incorporates the multi-task feature extraction algorithm studied in this paper,and realizes functions including administrator login and registration,vein information collection,and vein identity authentication.The system is built in the form of a website using the MVVM model.Both the front and back ends use a lightweight,highly customizable framework,which also makes the entire system more concise,and later expansion and maintenance are easier.