Research On Finger Vein Recognition Algorithm Based On FPGA And Convolutional Neural Network

Personal identity authentication technology is gaining increasing attention as society progresses,especially in the areas of security and consumption.Traditional identification methods have been unable to meet the needs of information security in modern society.Biometrics technology has received more and more attention from the society.As one of the most important biometrics technology,finger vein recognition technology has the characteristics of low cost,small acquisition device,non-contact collection and user friendliness compared with other biometric identification technologies.This paper first analyzes the basic theoretical knowledge of biometrics,image preprocessing,convolutional neural networks,FPGA acceleration calculation optimization,etc.,and then studies the three parts of FPGA on convolutional neural network accelerated computation optimization,image preprocessing and finger vein recognition system.The main work and research results of the thesis are summarized as follows:(1)FPGA optimization research on accelerating calculation of convolutional neural network: For convolutional neural networks,different optimization objects are divided into processor system and programmable logic,and the optimization method is studied.Implementing the inter-layer module multiplexing based on the pipeline structure on the processor system,and establishing a corresponding conflict processing mechanism to solve the signal conflict caused by multiplexing;using the segmentation parameters in the previous method as parameters in the programmable logic,using the HLS tool Design different layer accelerators.Finally,dynamic fixed-point operations are used instead of floating-point operations to reduce storage requirements and the resources consumed by storage transfers,while meeting the accuracy requirements of different network layers.(2)Pretreatment of finger vein images: mainly including region of interest extraction of finger vein,image normalization and data expansion.The finger vein quality problems that occurred due to individual differences and collection environment and equipment problems during the imaging process were analyzed.According to the clear edge feature of the finger and the high transmittance of the near-infrared light of the interphalangeal joint,the maximum intrinsic matrix method is used to extract the region of interest of the finger vein.Through the methods of gamma transformation,shear transformation,translation,rotation and partial amplification,a finger vein database with varied diversity is constructed.(3)Design of finger vein recognition system: The software and hardware framework of finger vein recognition system is given,and the system platform is built,including Linux system transplantation and project startup file compilation.The finger vein recognition system was verified on the ZCU102 development board.The data fusion database was used to integrate the finger vein database.The AlexNet network and the SqueezeNet network were used respectively to compare the accuracy of finger vein recognition and power consumption performance.Experiments have shown that the use of the system for finger vein recognition can improve the recognition efficiency of finger veins within an acceptable accuracy loss range.