在虹膜辨識系統中，segmentation是其中最為重要的一環，segmentation的品質好壞左右著虹膜辨識最終的成功率。在過去的研究中，已經開發出了許多的segmentation演算法，如neural network、Hough Transform，但是未曾出現過「評估」segmentation品質的演算法，所以也無法對segmentation的正確與否給予一個客觀化的指標。因此，我們開發出了一個方法叫KIRD，它可以針對segmentation的品質給出一個數值化的指標，可以在不需人工介入的情況下，正確的評估segmentation的好壞與否。並且，我們在KIRD的基礎上，開發出了一套叫作AILIS的segmentation演算法，它是一個會在迭代中學習、具有高度應變性的演算法。在每一輪迭代中，AILIS都會根據前一輪的結果自動的學習並優化機器學習模型，以此產生出品質更佳的segmentation。根據實驗結果，AILIS可以將ICE虹膜資料庫（灰階影像）中99.39%的眼部影像成功的生成品質極佳的segmentation，在UBIRIS虹膜資料庫（彩色影像）中也有94.60%的成功率，並且在後續的大規模虹膜辨識實驗也驗證了AILIS的有效性與高度適應性。;Iris segmentation is one of the most important pre-processing stage for an iris recognition system. The quality of iris segmentation results dictates the iris recognition performance. In the past, methods of either learning-based (for example, neural network) or non-learning-based (for example, Hough Transform) have been proposed to deal with this topic. However, there does not exist an objective and quantitative figure of merit in terms of quality assessment for iris segmentation (to judge whether a segmentation hypothesis is accurate or not). Most existing works evaluated their iris segmentation quality by human. In this work, we propose KIRD, a mechanism to fairly judge the correctness of iris segmentation hypotheses. On the foundation of KIRD, we propose AILIS, which is an adaptive and iterative learning method for iris segmentation. AILIS is able to learn from past experience and automatically build machine-learning models for iris segmentation for both gray-scale and colored iris images. Experimental results show that, without any prior training, AILIS can successfully perform iris segmentation on ICE (gray-scale images) and UBIRIS (colored) to the accuracy rate of 99.39% and 94.60%, respectively. Large-scale iris recognition experiments based on AILIS segmentation hypotheses also validated its effectiveness, compared to the state-of-the-art algorithm.
