基於三維灰階伴隨矩陣之虹膜辨識

在現今生活數位化的時代，人們為了追求更便利的生活方式與安全性問題，使得以生物辨識為基礎之身份認證技術為人們所重視。在多種不同生物辨識技術中，以虹膜影像較為穩定，運用在辨識上有良好的效果，故本論文將深入探討以人眼虹膜影像為基礎的虹膜辨識技術。本論文採用UBIRIS所提供的人眼虹膜影像資料庫，進行實驗測試。系統架構主要包含三個模組：影像前處理、特徵萃取以及分類辨識模組。輸入之人眼影像，由影像前處理模組利用影像處理演算法，自人眼影像中取得所需的虹膜影像。再經由特徵萃取模組以三維灰階伴隨矩陣（3D Gray Level Co-occurrence Matrix, 3D GLCM）計算虹膜特徵資訊，取得代表使用者身份的虹膜特徵碼（Iris code）。最後，分類辨識模組，則利用特徵碼進行比較分類及註冊的動作，來達到本論文所需要辨識的目的。以UBIRIS資料庫的241類不同人眼，共1,205張影像，對系統進行測試，影像前處理模組，能成功自人眼影像切割出虹膜區塊並強化的成功率，可達87.22%。本論文在特徵萃取方式，建立三維灰階伴隨矩陣，並以投影方式來統計像素組之間的關係，最後再以不同統計特徵組合下，以等錯誤率（equal error rate, EER）為標準，得到辨識成功率為98.87%（64 bytes）、99.42%（128 bytes）與99.65%（192 bytes）。甚至再要求系統之錯誤接受率（false acceptance rate, FAR）為0%的情況下，辨識成功率依然能保有90%（acceptance of authentic, AA）以上。最後，本論文將針對實驗的數據結果進行分析比較，來驗證建構本系統時所提出的相關推論，以供後續研究作為參考。

關鍵字

生物辨識；虹膜辨識；鑑別；灰階伴隨矩陣

並列摘要

With an increasing emphasis on security and convenience, personal authentication based on biometrics has been receiving extensive attention. Among many different biometric technologies, this thesis examines iris recognition technique for personal verification and develops a good performance recognition system based on human iris texture features. The proposed system consists of three modules: image preprocessing, feature extraction, and recognition modules. Image preprocessing module uses some image processing algorithms to localize the region of interest of iris from the input image. The feature extraction module adopts the 3D Gray Level Co-occurrence Matrix (3D GLCM) as the discriminating texture features. The system encodes the features to generate its iris feature codes. Finally, the system applies these feature codes for iris matching in recognition module. It is implemented and tested on the UBIRIS iris image database. The Experimental results show that the system has an encouraging performance on the UBIRIS database (including 1205 images from 241 classes). In the image preprocessing module, we checked the accuracy of the boundaries subjectively and obtained the success rate of 87.22%. Feature extraction method based on 3D GLCM use the projection to gather statistics of relation between pixel-tuples, and utilize the different feature combination, we attain the recognition rates up to EER = 99.5624% (64 bytes)、99.2308% (128 bytes)、99.9291% (192 bytes) and 99.9564% (256 bytes) (according to equal error rate, EER), respectively. Even under the circumstance of FAR = 0%, the system still approaches the recognition rates above 90% (acceptance of authentic, AA). This thesis analyzes the experiment results to verify the related inferences of the proposed system and provides useful information for further research.