【作者】 刘岩；

【导师】 于哲舟；

【作者基本信息】 吉林大学 ， 软件工程， 2021， 硕士

【摘要】 随着社会的发展水平逐步提高,社会信息化程度已经开始有了大幅度的提升,人们的日常生活各方面都与网络产生密切的联系,因而人们对自身隐私安全性的问题有了更加深刻的认知和理解,也对于身份安全认证系统的认证精确性提出了更高的要求。传统的身份认证识别方法存在着许多至今未曾解决的难题,不能很好的做到快速准确的自动身份认证,利用人体自身的生物特征信息进行身份认证是目前使用到的认可度最高的方式。在对各种生物特征信息的识别方法做了广泛对比研究后,虹膜身份认证识别系统因为其特征信息具有唯一性、不可侵犯性和可靠性的天然优势而受到各界学者广泛重视。对虹膜图像进行虹膜区域的定位是虹膜识别系统的关键性环节,虹膜定位的速率和准确性对身份识别认证的效果有直接的影响。在通过虹膜进行身份认证的全部步骤中,最关键的一步就是把虹膜定位出来,为了在之后的认证步骤中达到识别的精确性,必须在定位环节找到虹膜的确切边界,传统的虹膜定位方法由于客观原因不能够使得虹膜定位的准确率很高,神经网络算法具有极强的计算能力,不仅仅能够缩减处理图片的时间,还可以同时进行批处理的操作,特征提取模块具有较强的特征提取能力,能够相对精确的定位出虹膜的内外边界。对目标进行检测的主要目的是将图像中感兴趣的物体和不需要的背景割裂开,正是基于目标检测算法优良的性能和表现出的显著效果,所以在这篇文章中我们将YOLOV3算法作为基本框架以求在虹膜图像定位上有所突破。并在其基础上加以改进,使其在具有速度快的优势的同时也能够兼顾准确率高的优点。为了增强YOLOV3算法的特征提取能力,解决由原始特征提取模块算法所造成的退化问题,本文将特征提取模块换成了Densenet算法,并在其基础上通过复制骨干网络得到辅助网络使其更有利于检测小目标,为解决实验中虹膜图像样本少,容易造成的过拟合现象,本文加入了Dropblock正则化来解决,并且为解决因采集等因素导致的虹膜尺寸的不一致性、虹膜图像语义信息不够丰富的问题,加入了特征金字塔,用Non-local注意力机制来增强图片获取到的语义信息。本文模型的精确率和召回率与基础的YOLOV3算法相比较有了明显提高,与经典的定位虹膜的方式对比,本文的虹膜定位准确率也显著提升,经过对比本文的模型测试精确率为97.1%,它的综合性能占据绝对优势。更多还原

【Abstract】 With the development of society,the degree of social informatization has begun to improve greatly.Human life has a close contact with the internet.Therefore,people have a deeper understanding of the privacy security,and therefore,it puts forward higher demands for the authentication correctness of identity security authentication system.Traditional identification methods have many shortcomings that have not been solved so far.Therefore,it is not able to achieve rapid and exactness automatic identity authentication.It is the most recognized way to use biometric information of human body to authenticate.After a wide range of comparative studies on the recognition methods of biometric information,iris authentication system has been widely valued by scholars in all walks of life because of its unique,inviolable and reliable natural advantages.The iris region position of iris image is the key link of iris recognition system.The facilitate and exactness of iris position have direct influence on the effect of identity recognition and authentication.The main goal of this step is to supply excellent input for the next steps of feature information extraction and matching.In all the steps of identity authentication through iris,the most important step is to locate the iris.In order to identify the exactness in the subsequent authentication step,the exact boundary of iris must be found in the positioning link.Traditional iris position method can not make the iris position accuracy very high due to objective reasons,and neural network has a strong computing power,not only It can compress the operation time of one picture,and also can perform batch processing at the same time.The feature extraction module has strong feature extraction ability and can locate the inner and outer boundary of iris relatively accurately.The advocate intention of target find is to split the aims of interest and unnecessary downplay in the picture.Separating the aims and unnecessary parts in the picture is the main function of the target detection model.The target detection algorithm is so good in performance and remarkable effect.Therefore,we use YOLOV3 as the basic framework to make a breakthrough in iris image position.And on the basis of the improvement,it can have the advantages of high speed and high accuracy.so as to enhance the feature extraction ability of the feature extraction module of YOLOV3 algorithm and solve the degradation caused by the original feature extraction module algorithm,this paper replaces the feature extraction module with densene algorithm,and obtains the auxiliary network through the replication backbone network to make it more convenient to detect small targets.So as to decode the puzzle of the over appropriate caused by the small iris image samples in the experiment,the difficulty of over fitting is simple to cause So as to solve the inconsistency of iris size caused by collection and other factors,the semantic information of iris image is not rich enough.The feature pyramid is added.The non local attention mechanism is used to enhance the semantic information obtained by the image.The exactness and recall rate of the model are compared with the basic YOLOV3 algorithm compared with the classical method of iris location,the exactness of iris position is also greatly improved,and its comprehensive performance occupies an absolute advantage.更多还原