基于改进YOLO的虹膜快速定位检测算法  

Iris Fast Location Detection Algorithm Based on Improved YOLO

作　　者：周奥 杨岗[1] 闫磊 张东兴 ZHOU Ao;YANG Gang;YAN Lei;ZHANG Dong-xing(Southwest Jiaotong University,School of Mechanical Engineering,Chengdu 610031,China;R&T Center,CRRC Qingdao Sifang Co.,Ltd.,Qingdao 266111,China)

机构地区：[1]西南交通大学机械工程学院,成都610031 [2]中车青岛四方机车车辆股份有限公司技术中心,青岛266111

出　　处：《科学技术与工程》2022年第33期14801-14808,共8页Science Technology and Engineering

基　　金：国家重点研发计划(2020YFB1200300ZL);国家重大技术装备攻关工程项目(系列化中国标准地铁列车研制及试验);成都市重点研发支撑计划(2019-YF05-02685-SN)。

摘　　要：虹膜定位是虹膜识别系统中不可或缺的环节,针对传统的虹膜定位方法对镜面反射、眨眼等复杂环境下质量差的虹膜图像定位准确率低、计算复杂度高和鲁棒性差等问题,提出了一种基于改进YOLOv3模型的虹膜快速定位方法。针对眼周图像中虹膜内、外圆尺寸变化不大,将YOLOv3网络的多尺度结构改进为双尺度检测;引入了轻量级网络Mobilev3中bneck块来改进特征提取网络,减小模型复杂度;利用K-means++算法对虹膜数据集进行类聚,获得更优的锚点框;模型边框损失函数采用LossGIoU改进原均方差(mean squared error,MSE)损失函数;利用虹膜特有几何特征,将模型矩形预测框更改为圆形预测框。在CASIA-IrisV4数据集验证表明,改进模型定位准确率为96.32%,平均精度均值(mean average precision,mAP)为99.37%,检测速度为49.4帧/s,模型参数减少到4.13×10^(6)。结果表明改进后的模型较小,并且能够快速精准对虹膜区域定位,具有较高鲁棒性,能够满足虹膜实时定位的场景。Iris location is an indispensable link in the iris recognition system.Aiming at the problems of traditional iris location methods for poor-quality iris image positioning in complex environments such as specular reflection and blinking,low accuracy,high computational complexity,and poor robustness were proposed.Aiming at the small changes in the inner and outer circles of the iris in the periocular image,the multiscale detection of the YOLOv3 network structure was improved to dual-scale detection.An rapid iris location method based on the improved YOLOv3 model was proposed.The bneck block in the lightweight network Mobilev3 was introduced to improve the feature extraction network and reduce the complexity of the model.K-means++algorithm was used to cluster iris data sets to obtain better anchor frame.The improved mean square error(MSE)loss function was used as the frame loss function of the model.Based on the unique geometric features of iris,the rectangular prediction frame of the model was changed to a circular prediction frame.Validation on the CASIA-IrisV4 data set shows that the positioning accuracy of the improved model is 96.32%,the mean average precision(mAP)is 99.37%,the detection speed is 49.4 frame/s,and the model parameters are reduced to 4.13×10^(6).The results show that the improved model is small,and can quickly and accurately locate the iris area,has high robustness,and can meet the scene of iris real-time location.

关 键 词：虹膜定位 轻量级网络 YOLOv3 锚点框 损失函数 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]