Abstract:

In the iris recognition system, the segmentation of heterogeneous iris images（visible and infrared images） is the most important and challenging task. The difficulty of this task is that for heterogeneous iris images, both the accuracy and speed must be considered. This paper proposes a neural network model PI-Unet（Precise Iris Unet） suitable for heterogeneous iris segmentation, as well as data enhancement methods and loss functions for training the network model. The PI-Unet Encoder and Decoder are experimentally explored to obtain a network structure that can take into account both accuracy and speed. The data enhancement method and loss function proposed in this article are used to train the network. Using CASIA-iris -Intervel-v4 and UBIRIS.v2 iris image database to test the accuracy, parameter and calculation of the network. The test results show that the data enhancement method and loss function proposed in this paper can effectively improve the accuracy of heterogeneous iris segmentation. Compared with traditional iris segmentation algorithms and other iris segmentation neural networks, PI-Unet has better segmentation accuracy of heterogeneous iris images, less parameter and calculation, which can be applied to low-performance edge computing devices.

Key words: iris recognition, heterogeneous iris segmentation, neural network, data enhancement, loss function

摘要：

在虹膜识别系统中，异质虹膜图像（可见光和红外图像）的分割是最重要且最有挑战性的一个任务，该任务的难点在于针对异质虹膜图像，要同时兼顾虹膜分割的准确率和快速性。提出了适用于异质虹膜分割的神经网络模型PI-Unet（Precise Iris Unet）以及用于训练该网络模型的数据增强方法和损失函数。对PI-Unet的Encoder和Decoder进行实验探索，得出能同时兼顾准确率和快速性的网络结构，将提出的数据增强方法和损失函数用于该网络进行训练，在CASIA-iris-intervel-v4和UBIRIS.v2虹膜图像数据库上测试该网络的准确率、参数量和计算量。测试结果表明，提出的数据增强方法和损失函数能有效提高异质虹膜分割准确率，PI-Unet与传统虹膜分割算法和其他虹膜分割神经网络相比，对异质虹膜图像的分割准确率更高且参数量和计算量更少，能够适用于低性能的边缘计算设备。

关键词: 虹膜识别, 异质虹膜分割, 神经网络, 数据增强, 损失函数