Abstract: Iris recognition technology, with its excellent accuracy, security, and stability, has shown significant potential in border management, public security, financial payment and other fields. The current iris recognition system shows stable performance under the condition of constraints on user status and acquisition equipment, but it cannot adapt to the current complex and diverse open scenes. Open scenarios contain a large number of uncertain acquisition factors, for example, iris images acquired in open scenarios are easily interfered by factors such as eyelashes, hair blockage, and specular reflection, etc. These uncertainties often lead to an overall decline in image quality, resulting in a significant decline in the performance of iris image segmentation and feature extraction. Therefore, the research on iris recognition technology should further break through the existing limitations and expand from restricted scenarios to open scenarios with fewer constraints for research. In recent years, deep learning algorithms have been widely applied to iris recognition, aiming to improve the adaptability of the system to open scenarios. The current status of the application of deep learning technology in iris recognition is reviewed, and its key role in improving recognition accuracy in open scenarios is summarized. Firstly, the performance of various deep learning models in iris segmentation, iris feature extraction and feature matching tasks are analyzed and compared, and their advantages and limitations are expounded. Afterwards, the common iris datasets and their characteristics are systematically summarized. Lastly, new challenges and potential directions for future exploration of iris recognition are pointed out. This paper provides a comprehensive and in-depth understanding of deep learning-based iris recognition technology, which is conducive to promoting the innovation and optimization of iris recognition technology in practical applications.

Key words: iris recognition, biometrics, pattern recognition, computer vision, deep learning

摘要： 虹膜识别技术是一种十分安全的生物识别技术，以其卓越的精确性、安全性和稳定性等特点在边境管理、公共安全、金融支付等领域展现出巨大潜力。当前的虹膜识别系统在约束用户状态和采集设备的条件下展现出了较为稳定的性能，但是无法适应目前复杂多样的开放场景。开放场景中包含了大量不确定采集因素，例如采集的虹膜图像容易受到睫毛、头发遮挡和镜面反射等因素的干扰，这些不确定性因素往往会造成图像质量的整体下降，导致虹膜图像分割和特征提取环节性能的显著下降。因此，虹膜识别技术的研究应该进一步突破现有的限制，从受限制的场景扩展到约束条件更少的开放场景中进行研究。近年来，深度学习算法已被广泛应用于虹膜识别，旨在提升系统对开放场景的适应性。对深度学习技术在虹膜识别领域的应用现状进行了综述，总结了其在提高开放场景下识别精度的关键作用。首先，分析比较了各类深度学习模型在虹膜分割、特征提取和特征匹配任务中的表现，阐述了它们的优势和局限；接着，系统地总结了常见的虹膜数据集及其特性；最后，指出了虹膜识别任务新挑战及未来探索的潜在方向。提供了对基于深度学习的虹膜识别技术全面而深入的认识，有利于促进虹膜识别技术在实际应用中的创新和优化。

关键词: 虹膜识别, 生物特征识别, 模式识别, 计算机视觉, 深度学习