Research On Finger Vein Recognition By Fusion Of Vision Transformer And Capsule Neural Network

In recent years,finger vein recognition has been widely used in various identity authentication scenarios due to its advantages of high security and convenience.In the context of the global epidemic,finger vein recognition gradually tends to be unconstrained and non-contact.During the collection process,it is easily affected by factors such as finger position changes,surrounding environment,and lighting,resulting in a large inner-class gap in the collected finger vein images;In large-scale and multi-user finger vein recognition tasks,traditional finger vein recognition methods are gradually replaced by deep learning methods.However,the finger vein image has complex texture distribution,the distance between the near-end and the far-end finger vein texture is far away,and there are few samples of a single finger image,so it is difficult to effectively identify it.The above problems lead to poor recognition performance based on the deep learning model,which poses new challenges to the deep learning-based finger vein recognition technology.In view of the above problems,a novel fusion of vision transformer and capsule neural network finger vein recognition mode was proposed,it can combine the advantages of transformer in processing high-level vision information and the advantages of capsule neural network in processing detailed feature information,and better adapt to the transformation of tiny feature information in finger vein images.First of all,a linear embedding method for finger vein images is proposed,and the complete finger vein images are input into the model to extract texture feature information,so as to better capture the spatial information in finger vein images;Secondly,in order to be able to pay attention to the feature information of the finger vein image globally from a macro perspective,a vision transformer finger vein image processing module is proposed,which can better handle the long-distance dependencies in the finger vein image and capture the relationship between the pixels of the finger vein image,and realize the global perception of finger vein features,and improve the ability of the model to process texture features;In addition,insufficient single-class samples of finger veins can easily lead to the problem of model overfitting.A capsule network finger vein image processing module is proposed,which can capture information such as gesture,position,scaling and rotation of finger veins.This enables the model to effectively learn and extract richer vein feature information,alleviates the model’s demand for finger vein data samples,and improves the accuracy and stability of recognition.In this paper,performance experiments,ablation experiments,and comparative experiments were analyzed in four public finger vein data sets,and the deployment process of the model when it was applied to specific tasks of finger vein recognition was analyzed.The average recognition accuracy of the model on the four finger vein datasets is over 95%,and the average equal error rate is less than 1.5%.The experimental results prove that the model can extract global and local finger vein features,and realize high-performance finger vein recognition.