刘正琼,丁力,凌琳,李学飞,周文霞.基于字符编码与卷积神经网络的汉字识别[J].电子测量与仪器学报,2020,34(2):143-149 |
基于字符编码与卷积神经网络的汉字识别 |
Chinese character recognition based on convolutional neural network and character encoding |
|
DOI: |
中文关键词: 汉字识别 卷积神经网络 字符编码 过拟合 批标准化 |
英文关键词:Chinese character recognition convolutional neural networks character encoding overfitting batch normalization |
基金项目:安徽省科技攻关计划(1604a0902182)资助项目 |
|
|
摘要点击次数: 497 |
全文下载次数: 1 |
中文摘要: |
汉字识别是人工智能与模式识别领域中重要的研究内容,针对现有的研究仍然存在着参数调整难度大、训练样本数少、不能识别所有常用字符等问题,提出了一种基于字符编码与卷积神经网络的汉字识别方法,首先通过查询字库得到所有字符信息,以utf 8编码方式与多种字体编码文件进行编码输出字符图像,再进行多种图像处理后得到数据集,并利用深度卷积神经网络进行训练识别,在网络训练中通过数据扩增、批标准化、RMSProp优化等方式进行优化,同时加入正则化和Dropout防止过拟合。实验结果表明,所提方法对于汉字的识别率达到了9808%,与Alexnet、LeNet 5相比,使用同一数据集在识别准确率上提高了937%、2114%,实现了一个识别率高、特征提取能力与泛化能力强的神经网络。 |
英文摘要: |
Chinese character recognition is an important research content in the field of artificial intelligence and pattern recognition. Existing research still has problems such as difficulty in parameter adjustment, small number of training samples, and inability to identify all common characters. Aiming at these problems, we propose a Chinese character recognition method based on character encoding and convolutional neural network. First, we obtain all the character information by querying the font database, which are encoded and outputted by using UTF 8 encoding method and various font encoding files to generate character images. Further, we apply various of image processing to obtain the new character image dataset. Then, we propose a deep convolutional neural network for Chinese character recognition. In the training procedure, data augmentation, batch normalization, RMSProp optimization are optimized, regularization and dropout are used to prevent over fitting for optimization. The experimental results show that the proposed method is simple yet effective, the recognition accuracy rate for Chinese characters is 9808%. Compared with Alexnet and LeNet 5, we obtain a significant improvement by 937% and 2114%. A neural network with high recognition rate, strong feature extraction ability and generalization ability is realized. |
查看全文 查看/发表评论 下载PDF阅读器 |
|
|
|