【作者】 王超；

【导师】 刘元宁；

【作者基本信息】 吉林大学 ， 软件工程（专业学位）， 2022， 硕士

【摘要】 随着科技的进步,人们之间的信息交流越来越密切。在日常生活和行业中,都离不开信息的交互,因此人们对安全信息越来越重视。传统的认证方式不足以满足保护个人数据安全的要求。随着计算机技术的不断提高,虹膜身份识别技术凭借着防伪性、不可侵犯性以及更高的稳定性引起了国内外学者的关注。虹膜定位是虹膜识别系统中一个极其重要的步骤,这一步骤的目的是为了过滤掉无效的信息,定位出虹膜图像中内边界和外边界。目前的一些虹膜定位算法一般能够定位出虹膜的位置,但鲁棒性较差,而且传统的定位算法一次只能处理一张图片,计算量大、参数量大、定位时间慢,导致虹膜识别的整体性能也受到了很大的影响。近些年来,深度学习在人类视觉领域逐渐兴起,许多基于深度学习的目标检测算法可适用于各种场景,同时,这也为虹膜定位方法上提出一些新思路。因此本文将目标检测相关方法引入到虹膜定位算法中,目标检测算法可以将背景干扰部分以及我们的定位部分分开,而且能够准确的定位虹膜内外边界,并注明类别,每张图像的处理时间大大缩短,能够进行批处理操作。其中YOLO目标检测器的微调模型产生了高精度定位,克服了噪声、眼睑、睫毛和反射等问题,而且速度快。因此使用YOLOv4作为基本框架,并在此基础上进行改进,提出了一个新的模型Ghost-YOLOv4-CBAM。主要工作如下:(1)采用了中科院的CASIA-irisV4-Lamp数据集、吉林大学的JLU-6.0数据集以及JLU-7.0数据集,并使用Label Img工具进行手工标注数据集,并对实验数据集进行了大量的预处理工作,更好的适用于本文的模型。为了防止大尺寸先验框比小尺寸的先验框会产生更大的误差值,使用K-means算法对虹膜数据集重新进行聚类,并采用IOU距离来代替欧式距离作为度量距离,更准确的预测目标的位置。(2)为了解决YOLOv4算法中主干特征提取网络模型较大、计算资源消耗量大以及时间成本高等问题,将YOLOv4主干特征提取网络CSPDarknet53替换为更为轻便、计算复杂度更低的Ghost Net网络,并将Ghost Net网络的1×1卷积层、池化层以及全连接层去掉,保证与原特征提取网络具有相似的结构。(3)在此基础上,使用深度可分离卷积代替原模型PANet中的标准卷积,在未损失主干网络的特征能力的同时,减少了网络的参数,使模型更轻量化。由于YOLOv4算法中使用的数据集中小目标比较多,而虹膜数据集的大而清晰,所以将原始YOLOv4算法中的负责检测小目标的分支进行裁剪,保留中、大两个两个尺寸对虹膜进行定位,从而加快模型的推理速度。(4)为了让网络更加关注虹膜语义信息,增强有效特征和抑制无效特征,在融合特征信息之后以及特征融合网络中上采样过程加入CBAM注意力机制,最终得到的特征图输入到YOLO头部进行预测,从而提高虹膜检测的精确率。(5)通过在CASIA-irisV4-Lamp数据集和JLU-6.0数据集以及JLU-7.0数据集做大量对比实验和消融实验,其中对比实验主要是复现了SSD、Faster RCNN、YOLOv3以及Wilde算法。本文最终在三套数据集上得到的MAP值分别为0.9952,0.9941,0.9846,虹膜内外圆定位时间缩小为原来的三分之一,模型大小仅为39.14M。最终证明本文所提出的Ghost-YOLOv4-CBAM模型有效,在平均准确率基本上不损失的情况下,定位速度更快,模型更小。更多还原

【Abstract】 With the advancement of technology,people are exchanging information more and more closely with each other.The interaction of information is indispensable in everyday life and in industry,so people are placing more and more emphasis on secure information.Traditional authentication methods are not sufficient to meet the requirements of protecting personal data security.With the continuous improvement of computer technology,iris identification technology has attracted the attention of scholars at home and abroad by virtue of its anti-forgery,inviolability and higher stability.Iris location is an extremely important step in an iris recognition system.The purpose of this step is to filter out invalid information and locate the inner and outer boundaries of the iris image.Some current iris location algorithms are generally able to locate the iris,but are less robust.Moreover,traditional location algorithms can only process one image at a time,which is computationally intensive,has a large number of parameters and slow location time,resulting in a significant impact on the overall performance of iris recognition as well.In recent years,deep learning has emerged in the field of human vision,and many deep learning-based target detection algorithms can be applied to various scenarios,which also suggests some new ideas on iris location methods.This paper therefore introduces target detection-related methods into iris location algorithms,which can separate the background interference part from our location part,and can accurately locate the inner and outer borders of the iris and indicate the category,with much shorter processing time per image and the ability to perform batch processing operations.One of the fine-tuned models of the YOLO target detector produces highly accurate location,overcoming problems with noise,eyelids,eyelashes and reflections,and is fast.A new model,Ghost-YOLOv4-CBAM,is therefore proposed using YOLOv4 as the basic framework and improving on it.the main work is as follows.(1)The CASIA-irisV4-Lamp databases from CAS,the JLU-6.0 databases from Jilin University,and the JLU-7.0 databases were used,and the experimental databases was manually annotate the databases using the Label Img tool,with extensive Pre-processing work to better fit the model in this paper.In order to prevent large size a priori frames from generating larger error values than smaller size a priori frames,the iris databases was re-clustered using the K-means algorithm and the IOU distance was used instead of the Euclidean distance as the metric distance to predict the location of the target more accurately.(2)In order to solve the problems of large backbone feature extraction network model,high computational resource consumption and high time cost in the YOLOv4 algorithm,the YOLOv4 backbone feature extraction network CSPDarknet53 is replaced by the lighter and less computationally complex Ghost Net network,and the1×1 convolutional layer,pooling layer and fully connected layer of the Ghost Net network are removed.Guaranteed to have a similar structure to the original feature extraction network.(3)Based on this,the standard convolution in the original model PANet was replaced by using depth-separable convolution,which reduced the parameters of the network and made the model lighter without losing the feature capability of the backbone network.Since there are more small targets in the databases used in the YOLOv4 algorithm and the iris databases is large and clear,the branch responsible for detecting small targets in the original YOLOv4 algorithm is cropped to retain two sizes,medium and large,for iris location,thus speeding up the inference of the model.(4)To allow the network to focus more on iris semantic information,enhance effective features and suppress invalid features,a CBAM attention mechanism is added after fusing feature information and during Up-sampling in the feature fusion network,and the final obtained feature map is input to the YOLO head for prediction,thus improving the accuracy rate of iris detection.(5)By doing extensive comparison experiments and ablation experiments on the CASIA-irisV4-Lamp databases and the JLU-6.0 databases as well as the JLU-7.0databases,where the comparison experiments mainly replicated the SSD,Faster RCNN,YOLOv3 and Wilde algorithms.The final MAP values obtained on the three databases were 0.9952,0.9941,and 0.9846,respectively,and the iris inner and outer circle localization time was reduced to one-third of the original one,with a model size of only 39.14 M.The proposed Ghost-YOLOv4-CBAM model was finally proved to be effective,with essentially no loss in average accuracy,faster localization speed The model is smaller.更多还原