Scholarlink

The impact of pupil constriction and dilation on iris-based biometric systems has received proper attention in few works. This is due to the fact that in most current systems, information about pupil dilation is discarded when the iris region is normalized to a dimensionless polar coordinate system from which the iris code is obtained. However recent research has demonstrated the negative impact of pupil dilation on iris recognition performance. In this scenario we propose the use of data augmentation using automatic generated constriction and dilation of images from two well-known iris databases: CASIA-Lamp and CASIA-Thousand. To evaluate the proposed data augmentation we use the bounding boxes of non-normalized iris region as input to a fine-tuned Convolutional Neural Network (CNN) model (based either in ResNet-50 architecture), originally trained for face recognition. When compared to the state-of-the-art method that makes use of CNN, we observed that the proposed data augmentation consistently reduces the Equal Error Rate (EER) values by \(17\%\) and \(11\%\) respectively in the CASIA-Lamp and CASIA-Thousand databases when compared with the experiment without data augmentation.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Pupil Constrictions and Dilations Effects as Data Augmentation on an Iris Recognition CNN Approach

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Diego Rafael Lucio,
  • Luiz A. Zanlorensi,
  • Viviane Shiraishi Besson,
  • Yandre Maldonado e Gomes da Costa,
  • David Menotti

摘要

The impact of pupil constriction and dilation on iris-based biometric systems has received proper attention in few works. This is due to the fact that in most current systems, information about pupil dilation is discarded when the iris region is normalized to a dimensionless polar coordinate system from which the iris code is obtained. However recent research has demonstrated the negative impact of pupil dilation on iris recognition performance. In this scenario we propose the use of data augmentation using automatic generated constriction and dilation of images from two well-known iris databases: CASIA-Lamp and CASIA-Thousand. To evaluate the proposed data augmentation we use the bounding boxes of non-normalized iris region as input to a fine-tuned Convolutional Neural Network (CNN) model (based either in ResNet-50 architecture), originally trained for face recognition. When compared to the state-of-the-art method that makes use of CNN, we observed that the proposed data augmentation consistently reduces the Equal Error Rate (EER) values by \(17\%\) and \(11\%\) respectively in the CASIA-Lamp and CASIA-Thousand databases when compared with the experiment without data augmentation.