【作者】 夏伟；

【导师】 贾伟； 徐嘉文；

【作者基本信息】 合肥工业大学 ， 计算机技术（专业学位）， 2021， 硕士

【摘要】 在大数据驱动下的数字化社会中,个人的身份认证变得尤为重要,在此背景下,越来越多的身份认证方式应用在各个领域。掌纹识别作为一项新兴的生物特征识别技术,在过去的二十年里,多种传统掌纹识别方法被提出应用于相关场景中。近年来,深度学习的兴起为多项任务实现了新的突破,掌纹识别也逐渐在深度学习领域展开研究。然而,目前大多数深度掌纹方法往往只是简单的使用存在的经典神经网络完成识别任务,并没有充分的学习掌纹本质特征。因此,需要探索一个适用于掌纹特征的神经网络。此外,深度学习也是完成许多下游任务的重要利器。掌纹主线的提取一直是一个具有挑战性的工作,对深度掌纹主线的研究是一个新的尝试。本文的主要工作如下:(1)对当前较为典型的卷积神经网络(Convolutional Neural Network,CNN)和神经架构搜索(Neural Architecture Search,NAS)模型进行详细的阐述与分析。(2)利用迁移学习,在CNN模型和NAS模型上对掌纹识别进行了大规模的评估实验。选取17种卷积神经网络模型和20种神经架构搜索模型,在5个2D掌纹数据集和1个3D掌纹数据集上进行实验。我们针对不同的学习率进行实验测试,同时也和最先进的传统掌纹识别方法做了详细的比较,分析和讨论了迁移学习在掌纹识别上的应用前景。(3)在(2)的基础上,进一步提出一个深度掌纹融合网络TSPNet(Three-Stream Palmprint Network)进行掌纹识别。有了(2)中大规模评估实验结果作为指导,选择适合掌纹的分支骨干网络进行后续的融合识别。(4)提出轻量级分类神经网络EEPNet(Efficient and Effective Palm Net)进行掌纹识别。该网络具有速度快、精度高的优点,充分考虑了掌纹图像的本质特征,设计了两个全新的损失函数BL(Balanced Loss)和CL(Contrast Loss),并结合五个新颖的策略,达到了鲁棒的结果。(5)提出一个基于深度学习的掌纹主线提取与识别算法。该算法利用全卷积网络进行掌纹主线特征的提取,能够抑制掉很多的除主线之外的掌纹细纹,达到较为鲁棒的结果。此外,提出基于“点对点”和“段对段”的掌纹主线匹配方法,实验结果表明,掌纹主线识别技术能够取得令人满意的结果。更多还原

【Abstract】 In the digital society driven by big data,personal identity authentication becomes particularly important.Under this background,more and more identity authentication methods are applied in various fields.Palmprint recognition is a new biometric recognition technology.In the past twenty years,many traditional palmprint recognition methods have been proposed and applied in related scenes.In recent years,the rise of deep learning has enabled many tasks to achieve new breakthroughs,and palmprint recognition has gradually been researched in the field of deep learning.However,at present,most deep palmprint methods only use the existing classic network to complete the recognition task,and do not fully learn the essential characteristics of palmprint.Therefore,it is necessary to explore a neural network suitable for palmprint features.In addition,deep learning is also an important tool for accomplishing many downstream tasks.The extraction of palmprint principal line has always been a challenging task,and the study of deep palmprint principal line is a new attempt.The major works of this study are as follows:(1)The classical convolution neural network(CNN)and neural network architecture search(NAS)model are expounded and analyzed.(2)Using transfer learning,a large-scale evaluation experiment of palmprint recognition is carried out on CNN model and NAS model.Seventeen convolutional neural network models and twenty neural architecture search models were selected for experiments on five 2D palmprint databases and one 3D palmprint database.We do experimental for different learning rates,and compare them with the most advanced traditional palmprint recognition methods in detail,and analyze and discuss the application prospect of transfer learning in palmprint recognition.(3)On the basis of(2),a deep palmprint fusion network(TSPNet)is further proposed for palmprint recognition.With the results of the large-scale evaluation experiment in(2)as a guide,the branch backbone network suitable for palmprint is selected for subsequent fusion recognition.(4)A lightweight classification neural network(EEPNet)is proposed for palmprint recognition.This network has the advantages of high speed and high precision.Considering the essential features of palmprint image,two new loss functions,namely Balance Loss(BL)and Contrast Loss(CL),are designed,and five new strategies are combined to obtain robust results.(5)A deep palmprint principal line extraction and recognition algorithm is proposed.The algorithm uses the full convolution network to extract the feature of palmprint principal line,which can suppress many palmprint fine lines except principal line and achieve a more robust result.In addition,palmprint principal line matching method based on "point-to-point" and "segment-to-segment" are proposed.Experimental results show that palmprint principal line recognition technology can achieve satisfactory results.更多还原