|摘要: ||直至今日的人類社會，簽名認證系統(Signature Verification System)已被廣泛用於必須提供身份的各類場合中，包括公家機關以及一些商業行為上。以手寫簽名作為一種生物特徵擁有許多優點，它不像密碼會被人遺忘或是盜用; 也比起傳統生物認證系統來的節省成本，例如: 虹膜、人臉、指紋等。都需要較昂貴的硬體設備來截取影像資料; 而手寫簽名的取得，只需透過數位板或平板電腦即可，相較起來便宜。但手寫簽名認證也存在些缺點，它是一種困難的圖形識別問題，就算是同一個人，簽名樣本之間的變化程度很大。比起其它傳統生物認證方式，容易遭到有心人士模仿。|
本論文使用Leap Motion體感裝置擷取使用者簽名時的生物特徵，例如: 速度、角速度、用力程度等，並且對整個簽名軌跡分割成兩部份。分別屬於字體結構部分的實筆畫(Real Strokes)和提筆習慣部分的虛筆畫(Virtual Strokes)，本論文將兩者作為特徵抽取的依據。在特徵抽取方法中，依據實虛筆含有的生物特徵，製作出不同組合的特徵向量，並經過PCA(Principal Component Analysis)降維後，提升了鑑別能力。在實驗階段中，本論文針對在不同斷筆方法下抽取出來的特徵，使用倒傳遞類神經網路(Back Propagation Neural Network ,BPN)對樣本集進行訓練，將訓練出來的認證模型用於身份認證的應用上，本論文將會探討類神經網路在固定參數的情況下，不同特徵組合作用於身分認證的效果。實驗結果顯示本論文所提出的特徵抽取方法，應用在簽名確認系統上的確擁有較佳的效能，提升資料保密的安全程度。
;Signature Verification System has been widely used in several occasions that people need to provide their identities, including in government institution and in some business activities. Using handwritten signature as a biometric verification has lots of advantages, unlike passwords, passwords could be forgotten or embezzled. Also, it has lower cost than traditional biometric verification, such as iris, face, and finger print, which need expensive hardware equipment to capture images information. We can acquire handwritten signature just through a digitizer or a tablet, which is less expensive. However, there are still some drawbacks exist in handwritten signature. The changes in the characteristics of handwritten signature from the exact same person could be huge. Also, it is easier to be forged compared with other traditional biometric verification.
In this thesis, Leap Motion somatosensory device is used to detect the biometrics of users when signing their names, such as velocity, angular velocity, degree of force, etc. The device then divides the signature pattern into two parts, the shape of characters, Real Stroke, and the way individual uses its pen, Virtual Strokes. Those two parts are the basis of biometrics. According to the biometrics that real strokes or virtual strokes contain, it will produce different combination of features. The features are processed by PCA which promote the ability of identification. In the experiment, this paper aims at the biometrics produced from different ways of using pen, utilizing Back Propagation Neural Network, BPN, to train for sig-nature samples from different individuals and uses the model in identification. This paper will discuss the effect of identification under different parameters in BPN. The result indicates that applying different ways of using pen and detecting biometrics, both of which are mentioned in the paper, on signature confirmation system has better effect on identification, hence enhancing the security of information.