標題: 基於三軸加速度與陀螺儀之手寫簽名驗證系統A Handwritten Signature Verification System Based on G-sensor and Gyroscope 作者: 陳翌綺荊宇泰Chen, Yi-ChiChing, Yu-Tai資訊科學與工程研究所 關鍵字: 手寫簽名驗證;合成數據;支持向量機;特徵擷取;動態時間校正;直方統計;Handwritten Signature Verification;Synthetic data;Support Vector Machine;Feature Extraction;Dynamic Time Warping;Histograms 公開日期: 2017 摘要: 本篇論文利用感測器蒐集三軸加速度以及角速度數值，透過研究分析中種種處理數據的方法，實作簽名驗證，以達到身份驗證的目的。其中簽名實驗共有三種模式，包含本人長時間的簽名資料分析與測試、仿冒者透過不斷的練習偽造簽名、以及讓仿冒者直接將本人簽名之影像置於描圖紙之下仿冒之驗證，而其驗證之平均辨識率可達到93.47%。 研究方法中為每個身份建立一分類獨立模型以辨識身份的對錯。模型中使用的特徵包含握筆姿勢判斷、運筆力道及旋轉速度量值之機率分佈、以及其對於時關係之分析。而做資料比較時，必須先找出每個模型中本人資料的中心樣本，再將各個特徵分別以對應的距離公式(歐式距離、動態時間校正法) 計算建模資料與其之間的差距，得到之數據則為新型態之特徵表示，並再透過程式挑選部分資料用以製造仿冒簽名之資料特徵，最後利用支持向量機將其分為兩群，即為此簽名驗證系統中之分類模型。 此系統最大的特點在於，數據是加速度和角速度，此為無形之特徵，相較於影像更不易被竊取並透過練習而被仿冒成功。其二，每個身份皆為獨立的模型，所以不受資料庫內數據量多寡而影響辨識率。其三，改良特徵輸入模型的方式，讓資料間可以以多元的方式計算距離，而非傳統支持向量機皆以歐式距離之方式比較差距。In this thesis, through a variety of signal processing methods in research and analysis, accelerometer and gyroscope were used to implement a signature recognition system at the aim of identity authentication. Classified by recognition of selfsignature and counterfeit signatures, three modes are considered, including longtime analysis and testing of self-signature by identity, forged signature by counterfeiters with continuous practicing or directly through tracing paper. The average recognition rate can reach up to 93.47%. In the presented method, an independent model is established for each identity using their signatures to recognize it. The features of the signature used in this model include the gesture of the identity, the probability distribution of strength and speed of rotation during writing, and the values in time series. These features are considered as a point in high dimension space. A center is determined while the other points are used to establish a SVM as the independent model for the specific identity. The greatest benefit is that since these data collected from Koala are acceleration and angular velocity, which are invisible and difficult to imitate. In addition, each identity builds an independent model, therefore, the accuracy would not be affected by the amount of data in database. Last but not least, the input data that put into the model is based on distance, which can be calculated in a variety of way, such as using Euclidean distance and dynamic time warping on the features simultaneously. URI: http://etd.lib.nctu.edu.tw/cdrfb3/record/nctu/#GT070456017http://hdl.handle.net/11536/141445 Appears in Collections: Thesis