Research On Iris Recognition Method Based On Lightweight Neural Network

In the past two years,due to the epidemic,people need to register and identify themselves when entering and leaving public places.The currently commonly used face recognition system is vulnerable to mask obscuration.In this paper,we explore an iris recognition method for mobile from the field of iris recognition.Compared with traditional iris recognition methods,deep learning-based iris recognition methods have higher accuracy and system robustness,but their large number of participants and high computing and storage requirements make them difficult to deploy on mobile devices with limited hardware.To solve the above problems,this paper proposes a lightweight neural network combining lightweight network structure and model quantization techniques for iris recognition on mobile,and the specific work of the paper is as follows.First,this paper introduces the current status of domestic and international research on traditional iris recognition methods and neural network-based iris recognition methods,and describes the neural network lightweighting techniques that reduce the complexity of the latter models and their research status.Subsequently,this paper introduces the system architecture of iris recognition systems and convolutional neural networks,and investigates several classical lightweight network structures and model quantization methods.In order to solve the problem that it is difficult to run a large number of neural network models in mobile iris recognition systems,this paper proposes a lightweight neural network structure based on depthwise separable convolution structure for iris recognition.This method is optimized to address the accuracy degradation problem of lightweight network structure by combining cross-layer transmission mechanism,channel shuffle and dense connection techniques;and a dynamic truncation-based model quantization scheme is proposed to make the model more suitable for mobile deployment.Finally,in order to analyze the feasibility of the proposed method on mobile devices,the trained network models are ported to Android mobile devices for testing,and existing iris recognition networks and model quantization schemes are selected for comparison with the proposed lightweight neural network.