摘要
基于卷积神经网络能够直接从训练样本中提取特征并且具备权值共享等优势,本文提出了利用两级卷积神经网络实现纸币冠字号的识别方法。在字符分割过程中,考虑到待识别对象因破损、脏污等情况而引起的问题,提出了窗口移动配准法。实验表明,识别率可达99.99%以上,识别时间能控制在5 ms以内。
Based on the advantages that convolution neural network can directly extract characters from the training sample and weight sharing,a method of using the two levels CNN is proposed to recognize the RMB serial number. In the process of character segmentation,the method of window moving is proposed,considering the problems caused by the damage and smirch of the object to be identified. The experiment shows that the recognition accuracy can reach at least 99.99% and the recognition time can be limited to less than 5 ms.
出处
《辽宁科技大学学报》
CAS
2017年第2期133-137,共5页
Journal of University of Science and Technology Liaoning
关键词
卷积神经网络
权值共享
冠字号
convolutional neural network
weight sharing
serial number