生活有許多重要場合需要使用手寫簽名，當中牽涉龐大的利益，偽造簽名的問題也層出不窮。因此，建立快速、有效、科學化的簽名真偽辨識工具一直是個實際且重要的議題。現實生活中，收集專業的偽造簽名十分困難，因此根據經濟考量下提出二種不同的實驗，提供給使用者參考。實驗一是訓練時與最終測試時使用相同作家的簽名，實驗二是訓練時與測試時使用不同作家的簽名。一般來說，國際上常見的相關文獻以實驗二居多。深度學習中的卷積類神經網路 (convolutional neural network, CNN) 是近年興起的影像辨識方法，以其具有較高的模型績效而聞名。本研究使用ResNetv2，試驗了50、101、152 層對於2 種實驗的結果，每種組合收集了30 次的結果。結果方面，本研究發現實驗一的平均錯誤率只有0.02655 至0.0365 之間，實驗二與其他研究的績效比較之下，有比較低的錯誤驗收率。本研究發現在3種層數下，分類錯誤率的差異並不明顯。因此，在實務中使用時，似乎可以尋找更少層數的模型，來減少時間與硬體的耗用。另外，本研究發現到，整體而言，模型有錯誤驗收率低於錯誤拒絕率的傾向，這表示本研究所提出的CNN辨識系統在無法接受偽造簽名的簽名識別下具有優勢。

Handwritten signatures are required on many important occasions in life. As signatures are immensely related to human interests, issued with forged signatures constantly arise. It is, therefore, crucial to establish a fast, effective and scientific signature verification method. Most existing methods used to verify Chinese signatures have failed to achieve satisfying results. In contrast, the convolutional neural network (CNN) in deep learning that has emerged in recent years is a method of image recognition and is known for its high model performance. This study thus hopes to achieve good results in Chinese signature recognition by using CNN in deep learning. In practice, it is very difficult to collect professionally forged signatures. Therefore, based on economic concerns, two different experiments were proposed to test the performance of the CNN. In the first experiment, the signatures of the same owners were used during the training and the final test, while in the second experiment, the signatures of different owners are used. It is worth mentioning that in most of the existing documents the second method of experiment was adopted. In this study, the 50-layer, 101-layer, and 152-layer deep residual convolutional neural networks (ResNet-v2) were used respectively to establish the signature recognition system. Since there are two types of experiments, each with three types of network configurations, there are a total of six experiment combinations, and thirty replications for each combination have been conducted. Therefore, a total of 180 network performance data have been collected and analyzed. The results showed an error rate of the first experiment as between 0.026-0.0365. When we compared the performance of the second experiment with that in other studies, we found that the proposed method had a lower false acceptance rate. We also found that the classification error rates of the three types of network configuration showed no significant difference. This result implies that in reality the user can use fewer layers of network to reduce time and hardware consumption. Finally, in this research the false acceptance rate is lower than the false rejection rate in the six experimental combinations. It implies that the proposed CNN recognition system has the advantage of not accepting forged signature recognition.