由於每個人的指紋有著獨一無二的特性，因此在生物特徵辨識的領域中，不論是做確認或者是辨認，指紋一直都是最熱門的技術之一。為了使龐大的指紋資料庫的管理更為便利，並提昇指紋辨認的效率，我們通常會先將指紋分類成弧型 (arch)、帳型 (tented arch)、左箕型 (left loop)、右箕型 (right loop) 以及螺旋型 (whorl) 五類。目前已有許多種指紋分類的方法，這些方法各有其優缺點。本論文我們提出了一個了新的指紋分類系統，希望能只萃取少量的特徵而達到儘可能高的分類效果。在此系統中，首先利用一種有效的方法來將指紋影像轉成區域方向影像 (block directional image)，接著再透過一個註冊點 (registration point) 偵測的處理步驟，將每張指紋影像都定位於指紋的中心點，以此註冊點為中心取一適當的區域，從此區域中萃取出合適的特徵。最後藉由多維矩形複合式類神經網路(Hyper Rectangular Composite Neural Networks) 進行分類處理。我們使用 NIST-4 資料庫來訓練、測試我們的系統。 Fingerprints are one of the most popular biometrics techniques in both of verification and identification mode because the fingerprints of an individual are unique. To facilitate the management of large fingerprint database and to speedup the process of fingerprint identification, we will first classify fingerprints into several categories such as arch, tented arch, left loop, right loop, and whorl. Several different approaches have been proposed for fingerprint classification. Each has its own advantages and limitations. In this thesis, a new fingerprint classification system is introduced. The proposed system tries to use feature as fewer as possible, while to achieve correct recognition as high as possible. In this system, we first propose an efficient method to transform fingerprint image into block directional image. Then a registration point detection method is applied to locate the center of each block directional image. In the following, several feature are extracted from a window whose center is located at the detected registration point. Finally, a class of Hyper Rectangular Composite Neural Networks (HRCNNs) is trained for fingerprint classification. The system was tested on 4000 images in the NIST-4 database.
