虹膜辨識是一種以擷取虹膜特徵並加以比對以得出該虹膜所屬者身分的生物辨識方法。然而，虹膜影像的品質會對比對的結果有著各式的影響。如果不將品質不佳的影像排除，虹膜特徵的比對結果會受到影像模糊、反光點、眼皮等等影響而導致虹膜辨識的結果發生誤判。因為使用者經常會在實時虹膜辨識系統中提供不盡理想的虹膜影像或是照明設備曝光不足、過度曝光，誤判的情形將會更為嚴重。 本研究為了解決虹膜影像品質過差造成誤判的影響，測試了多種虹膜影像品質檢定方式並挑選出其中多個合適的檢定方式以在合理的時間內篩選出品質良好的影像。為了使多種虹膜影像品質檢定方式都能夠同時發揮功效，本研究嘗試了相加式品質計算、相乘式品質計算以及加權式品質計算於中國科學院虹膜資料庫(CASIA-Iris-Thousand V4)之中，並以相等錯誤率(Equal Error Rate)作為虹膜影像品質計算方式表現的評斷標準。本研究最終以加權過後的相乘式品質計算排除一半低於門檻值的影響後將相等錯誤率由3.1244%降至0.9092%。 ;Iris recognition is one of a biometric recognition method in which we extract iris features and compare them with other irises to identify whom the iris belong to. However, the performance of iris recognition is highly related to the quality of iris images. If images with poor quality are not excluded, the false non-matching of iris may happen because of images blur, reflections, or eyelids occlusion and the like. The problem becomes more evident in a real-time iris recognition system where users often provide non-ideal iris images, or when the iris images are underexposed or overexposed. To tackle this problem, in this dissertation we experimented multiple iris quality assessment algorithms and proposed feasible methods to select images with good quality in a reasonable time frame. Furthermore, large-scaled experiments are performed using CASIA-Iris-Thousand V4 database in order to figure out the optimized method for score fusion, for the goal of minimizing the equal error rate (EER) in the iris matching experiments. After the experiment, we used one of the best method (Exponentially Weighted quality assessment algorithm) to filter out half of the images. After such iris image quality thresholding, we are able to lower the EER of the large-scale iris recognition (with remaining iris images) from 3.1244% to 0.9092%, which proves the effectiveness of the proposed iris image quality assessment algorithm.
