IEEE Trans Pattern Anal Mach Intell. 2011 Sep;33(9):1744-57. doi: 10.1109/TPAMI.2010.227. Epub 2010 Dec 23.
Iris recognition typically involves three steps, namely, iris image preprocessing, feature extraction, and feature matching. The first two steps of iris recognition have been well studied, but the last step is less addressed. Each human iris has its unique visual pattern and local image features also vary from region to region, which leads to significant differences in robustness and distinctiveness among the feature codes derived from different iris regions. However, most state-of-the-art iris recognition methods use a uniform matching strategy, where features extracted from different regions of the same person or the same region for different individuals are considered to be equally important. This paper proposes a personalized iris matching strategy using a class-specific weight map learned from the training images of the same iris class. The weight map can be updated online during the iris recognition procedure when the successfully recognized iris images are regarded as the new training data. The weight map reflects the robustness of an encoding algorithm on different iris regions by assigning an appropriate weight to each feature code for iris matching. Such a weight map trained by sufficient iris templates is convergent and robust against various noise. Extensive and comprehensive experiments demonstrate that the proposed personalized iris matching strategy achieves much better iris recognition performance than uniform strategies, especially for poor quality iris images.
虹膜识别通常包括三个步骤，即虹膜图像预处理、特征提取和特征匹配。虹膜识别的前两个步骤已经得到了很好的研究，但是最后一个步骤涉及较少。每个人的虹膜都有其独特的视觉模式，局部图像特征也因区域而异，这导致了不同虹膜区域提取的特征码在稳健性和独特性方面存在显著差异。然而，大多数最先进的虹膜识别方法使用统一的匹配策略，即认为来自同一个人不同区域或不同人同一区域的特征是同等重要的。本文提出了一种个性化的虹膜匹配策略，使用从同一虹膜类别的训练图像中学习到的类特定权重图。当成功识别的虹膜图像被视为新的训练数据时，权重图可以在虹膜识别过程中在线更新。权重图通过为虹膜匹配的每个特征码分配适当的权重来反映编码算法在不同虹膜区域的稳健性。通过足够多的虹膜模板训练得到的权重图是收敛的，并且对各种噪声具有鲁棒性。广泛而全面的实验表明，所提出的个性化虹膜匹配策略比统一策略具有更好的虹膜识别性能，尤其是对于质量较差的虹膜图像。
