计算机工程与应用 ›› 2023, Vol. 59 ›› Issue (2): 120-128.DOI: 10.3778/j.issn.1002-8331.2106-0460

• 模式识别与人工智能 • 上一篇    下一篇

IrisBeautyDet:虹膜定位和美瞳检测网络

陈旭旗,沈文忠   

  1. 上海电力大学 电子与信息工程学院,上海 201306
  • 出版日期:2023-01-15 发布日期:2023-01-15

IrisBeautyDet:Neural Network for Iris Localization and Cosmetic Contact Lens Detection

CHEN Xuqi, SHEN Wenzhong   

  1. School of Electronic and Information Engineering, Shanghai University of Electric Power, Shanghai 201306, China
  • Online:2023-01-15 Published:2023-01-15

摘要: 虹膜活体检测是虹膜识别中涉及安全的重要环节之一,也是虹膜识别领域亟待解决的问题之一,其中美瞳检测是虹膜活体检测中最具挑战性的领域。提出了一种基于SSD(single shot multibox detector)目标检测网络的虹膜定位和美瞳检测算法IrisBeautyDet,并对网络结构进行轻量化处理,引入MobileNet主干网络显著减少模型计算量,极大提高速度。采用空间注意力和通道注意力机制,进一步提高模型准确率。实验表明,在CASIA-Iris和圣母大学NDCLD的活体和美瞳虹膜数据集上,该算法具有较好的泛化能力和鲁棒性,相比原始SSD算法,IrisBeautyDet具有更少的参数量、更快的实时性和更高的准确率。相比原始SSD网络模型,该模型大小从91.1 MB下降到26.1 MB,同时将检测速度从29.68 frame/s提高到41.88 frame/s,对活体类和美瞳类的检测精确率达到99.21%和98.61%。利用导向反向传播(guided-backpropagation)对检测特征图进行可视化,分析并优化网络模型使其更有效地提取美瞳纹理特征。

关键词: 美瞳检测, 虹膜活体检测, 呈现攻击检测, 注意力机制, 轻量级网络, 特征图可视化

Abstract: Iris liveness detection is one of the important aspects involving security in iris recognition and one of the urgent problems in the field of iris recognition, where cosmetic contact lens detection is the most challenging area of iris liveness detection. This paper proposes a cosmetic contact lens detection algorithm IrisBeautyDet based on SSD(single shot multibox detector) object detection network, and lightens the network structure to significantly reduce the computation cost and substantially improve the speed by using MobileNet. Spatial attention and channel attention mechanisms are effective for the model to considerably improve accuracy. Experiments indicate that the algorithm has better generalization ability and robustness on the CASIA-Iris dataset and the University of Notre Dame NDCLD dataset of liveness and cosmetic contact lens iris data. Compared with the original SSD algorithm, IrisBeautyDet has fewer parameters and more prominent real-time performance. And the size of this model decreases from 91.1 MB to 26.1 MB, while increasing the detection speed from 29.68 frame/s to 41.88 frame/s. And the model’s classification precision achieves 99.21% and 98.61% for the live category and the cosmetic contact lens category. Guided-backpropagation is used to visualize the detection of feature map, analyze and optimize the model to make it more effective in extracting texture features of cosmetic contact lens.

Key words: cosmetic contact lens detection, iris liveness detection, presentation attack detection, attention mechanism, lightweight network, feature map visualization