Abstract:

In view of the current human eye localization algorithm to deal with the complex environment of the anti-interference ability is not strong, the positioning accuracy is poor and there is no left and right eye classification, an algorithm for human eye location and left and right eye classification of iris image based on lightweight network is proposed. The EL-YOLO model is designed by using YOLO algorithm combined with the high-performance lightweight network model. The Generalized Intersection-over-Union（GIoU） is introduced into the loss function, which enables the network training to converge quickly and achieve high positioning accuracy. The experimental results on the datasets CASIA-IrisV4, MIR2016, and the datasets SEPAD_V1 and SEPAD_V2 collected by our lab show that the EL-YOLO model is small, runs fast, has a high accuracy of positioning and classification, and has a strong generalization ability.

Key words: iris recognition, human eye location, lightweight network, generalized intersection-over-union

摘要：

针对当前的人眼定位算法应对复杂环境的抗干扰能力不强、定位准确度较差以及无左右眼分类的问题，提出了一种基于轻量级网络的虹膜图像人眼定位及左右眼分类算法。利用YOLO算法结合高性能的轻量级网络模型设计EL-YOLO模型，损失函数引入广义交并比（GIoU），使得网络训练可以快速收敛，且定位精度高。在CASIA-IrisV4、MIR2016以及本实验室采集的数据集SEPAD_V1和SEPAD_V2上的实验结果表明，EL-YOLO模型较小，运行速度快，且拥有较高的定位及分类准确率，具有较强的泛化能力。

关键词: 虹膜识别, 人眼定位, 轻量级网络, 广义交并比