A Deep Learning-Based Offline Signature Verification Method by Single Known Sample

基於深度學習及單一已知簽名樣本的離線驗證方法

高信雄(Hsin-Hsiung Kao) ； 溫哲彥(Che-Yen Wen)

《中正嶺學報》 49卷1期 (2020/05) Pp. 23-34

https://doi.org/10.30188/JCCIT.202005_49(1).0003

摘要

Signature verification is one of the popular biometric techniques for personal identification. Although automatic signature verification systems have attracted researchers' attention for a long time, there are few attempts to perform the verification based on a single known sample. In this paper, we propose an off-line handwritten signature verification method by using a unique local feature extraction approach and deep convolutional neural network (CNN). In the training process, our CNN is trained by only single genuine (known) and some forged signature samples. In the testing process, the proposed method can verify a questioned signature as genuine or forgery (all questioned signatures and forged authors were not present in the training process). We use the open source dataset, Document Analysis and Recognition (ICDAR) 2011 SigComp in the experiments, and get the accuracy of 98.41%, FAR of 0.91% and FRR of 2.88% in our testing dataset.

關鍵字

handwritten signature verification ； convolutional neural network (CNN) ； deep learning ； forensic science

並列摘要

簽名驗證為主流的生物特徵識別技術之一，儘管相關研究已行之有年，卻鮮少有學者基於單一已知簽名樣本的驗證問題進行研究。本文以獨特的局部特徵擷取法結合深度卷積神經網路（CNN）提出創新的離線簽名驗證方法。本研究透過單一真實簽名樣本輔以數個偽造簽名進深度學習訓練，於測試過程中，所有未知簽名及偽造簽名的作者皆不曾出現在訓練過程，以確保系統的有效性。本文簽名樣本採自ICDAR 2011 SigComp公開資料集，並取得98.41%的準確率、0.69%的錯誤接受率（FAR）及2.88%的錯誤拒絕率（FRR）。

並列關鍵字

手寫簽名驗證 ； 卷積神經網路（CNN） ； 深度學習 ； 鑑識科學