Research On Finger Vein Recognition Algorithm Base On Artificial Neural Network

Identity recognition have become more and more necessary in information society,making biometric identification technology widely concerned.Thanks to many advantages such as uniqueness,anti-spoofing,and unforgeability,finger vein recognition technology undoubtedly has great research and use value.Among various recognition algorithms,artificial neural networks stand out by virtue of their powerful feature extraction capabilities and super high generalization.Therefore,combining the two effectively becomes a meaningful and promising job.This paper innovatively uses deep learning for finger vein recognition,overcoming the various drawbacks of traditional methods,and proposes several valuable solutions to the current difficulties.The main works are as follows:Firstly,we analyze the current situation and research background of finger vein recognition,and the complete finger vein recognition process is systematically explained.Secondly,aiming at the problem of poor quality of finger vein images and loss of image details due to improper exposure,a multi-branch and end-to-end convolutional neural network is proposed.Based on the attention mechanism and dilated convolution,an image reconstruction algorithm similar to high dynamic range is realized,which restores image details well and improves image contrast.Thirdly,for the problem of insufficient finger vein data to train deep learning models,a transfer learning method is proposed.Through the "pre-training-fine-tuning" mode,detailed experiments have been done on five models of different structures.Experiments show that the method proposed in this paper can successfully train a large model on a small data set,shorten the training time and make full use of the learned knowledge of the pre-training model,and has good generalization.In addition,in order to solve the problem of over-fitting of the convolutional neural network caused by insufficient finger vein samples,the graph neural network is used to finger vein recognition.This method has high recognition accuracy,reduces the algorithm's requirements for image quality,and improves the matching efficiency.Finally,neural networks are memory-intensive and computationally,which makes it difficult to deploy them on the embedded devices,in view of this,a model compression method is proposed.Through model pruning and model quantification,without affecting the recognition accuracy,the model's requirements for model size,inference time,power consumption,memory usage and other indicators are reduced.