Application of Convolutional Neural Networks in Offline Chinese Signature Recognition

• Main Authors: HUNG, WEI-CHIA, 洪偉嘉
• Other Authors: Hou, Tung-Hus
• Format: Others
• Language: zh-TW
• Published: 2018
• Online Access: http://ndltd.ncl.edu.tw/handle/97ra2d

碩士 === 國立雲林科技大學 === 工業工程與管理系 === 106 === 　　There are many important occasions in life that people need to handwriting signature. Because signature involved a huge people interest, fake signature emerged endlessly. Therefore it is a practical and important issue to establish a fast, effective and scientific signature verification method. Many tools have not been able to achieve good results in Chinese signatures. The convolutional neural network in deep learning is a method of image recognition that has emerged in recent years and is known for its high model performance. This study hopes to achieve good results in Chinese signature recognition by using convolutional neural network (CNN) in deep learning. 　　In real practical world, it is very difficult to collect professional forged signatures. Therefore, according to economic considerations, two different experiments are proposed to test the performance of the CNN. The first experiment used the signature of the same authors during the training and the final test. The second experiment used the signatures of different writers during training and testing. In general, most of the international literatures did the second experiment. 　　In this study, the 50-layer, 101-layer, and 152-layer deep residual convolutional neural networks (ResNet-v2) were used to establish the signature recognition system. Since there are two types of experiment and each with three types of network configurations, therefore there are six experimental combinations. Thirty replications for each combination were conducted, therefore, a total of 180 network performance data were collected and analyzed. 　　In the results, this study found that the performance of the first experiment was better than that of the second experiment. A comparison result of the performance of the second experiment with other studies was reported in this study. This study also investigated whether the large number of model layers will be over-fitted. It is found that the classification error rates of the three types of network configuration are not significant difference. This result implies that when used in practice the user can use a fewer layers of networks to reduce the time and hardware consumption. Finally, this research found that 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.