基于深度残差网络的文字识别算法研究  

Research on Text Recognition Algorithm Based on Depth Residual Network

作　　者：孟彩霞[1] 王腾飞 王鑫[1] MENG Caixia;WANG Tengfei;WANG Xin(Xi'an University of Posts and Telecommunications,Xi'an 710000)

机构地区：[1]西安邮电大学

出　　处：《计算机与数字工程》2019年第6期1487-1490,1501,共5页Computer & Digital Engineering

摘　　要：房产证、名片、公司营业执照等信息的收集通常是通过人工进行录入,而人工的录入方式存在效率低和容易出错的问题,因此汉字和数字的识别技术具有很重要的实用价值。由于汉字的种类繁多,传统的印刷体汉字识别存在精度低,不能满足实际应用需求等不足。随着深度学习的快速发展,多分类问题的精度得到显著改善,中文汉字识别得到了快速发展。但是针对汉字的识别深度学习网络存在网络模型复杂、容易过拟合的不足。若简化网络模型又会出现识别精度不高的问题。针对这些不足,基于深度残差网络(ResNet),提出一种改进的中文文字识别深度残差网络模型。实验结果表明,针对一级汉字,训练正确率达到98.7%,测试正确率达到97.6%。该方法识别精度高并且网络容易收敛,具有很高的应用价值。Property certificate,name card,business license and other information collection is usually done by manual input,manual entry way has low efficiency and error prone,so Chinese characters and digital recognition technology have very important practical value. Due to the variety of Chinese characters,the traditional printed Chinese character recognition has low accuracy and can not meet the needs of practical application. With the rapid development of deep learning,the accuracy of the multi classification problem has been significantly improved,and the Chinese character recognition has been rapidly developed. However,for the recognition of Chinese characters,the deep learning network has the shortcomings of complex network model and easy over fitting. If the network model is simplified,the recognition accuracy is not high enough. Aiming at these shortcomings,an improved residual network model of Chinese character recognition based on depth residual network(ResNet)is proposed. The experimental results show that the correct rate of training is 98.7%,and the correct rate is 97.6%. The method has high recognition accuracy and easy convergence,so it has high application value.

关 键 词：卷积神经网络 深度残差网络 光学字符识别 文字识别 

分 类 号：TP391.43[自动化与计算机技术—计算机应用技术]