计算机工程与应用 ›› 2022, Vol. 58 ›› Issue (21): 197-204.DOI: 10.3778/j.issn.1002-8331.2104-0053

• 模式识别与人工智能 • 上一篇    下一篇

基于多任务学习的可见光与近红外虹膜融合研究

尤轩昂,赵鹏,慕晓冬,朱永清,沈丹瑶   

  1. 中国人民解放军火箭军工程大学 作战保障学院,西安 710025
  • 出版日期:2022-11-01 发布日期:2022-11-01

Research on Fusion of Visible and Near-Infrared Iris Based on Multi-Tasking Learning

YOU Xuan’ang, ZHAO Peng, MU Xiaodong, ZHU Yongqing, SHEN Danyao   

  1. College of Operational Support, Rocket Force University of Engineering, Xi’an 710025, China
  • Online:2022-11-01 Published:2022-11-01

摘要: 针对在可见光虹膜识别中存在的虹膜纹理特征不明显与反射光斑问题,提出一种基于多任务学习的可见光与近红外无监督融合模型(MTIris-Fusion)。设计了基于改进DenseU-Net的端到端融合骨干网。设计了自适应权衡源图像重要信息保留度的损失函数,自适应保持融合结果与源图像之间的相似度,达到无监督的目的。通过弹性权重巩固(EWC)机制更新多融合任务的权重,避免了多任务网络中的灾难性遗忘。在PolyU Cross Spectral Iris数据集上的实验表明,与其他方法相比,该方法兼顾可见光虹膜的颜色纹理与近红外图像的结构信息并有效抑制了可见光图像中的光斑噪声,在虹膜图像质量增强领域具有重要应用价值。

关键词: 多任务学习, 虹膜图像融合, 可见光, 近红外, 损失函数

Abstract: Aiming at the problems in the visible light iris recognition that the iris texture features are not obvious and the reflected light spots, This paper proposes an unsupervised fusion model of visible light and near-infrared based on multi-task learning(MTIris-Fusion). The method designs an end-to-end fusion backbone network based on improved DenseU-Net. A loss function is designed to adaptively weigh the retention of important information of the source image, and adaptively maintain the similarity between the fusion result and the source image, so as to achieve the goal of unsupervised. Subsequently, the weights of the multi-converged tasks are updated through the elastic weight consolidation mechanism to avoid catastrophic forgetting in the multi-task network. Experiments on the PolyU Cross Spectral Iris data set show that compared with other methods, this method takes into account the color texture of the visible light iris and the structure information of the near-infrared image, and effectively suppresses the spot noise in the visible light image. It has important application value in the field of iris image quality enhancement.

Key words: multi-tasking learning, iris image fusion, visible light, near-infrared, loss function