基于改进卷积神经网络的身份证信息识别  

ID card information recognition with improved convolutional neural network

作　　者：高尚 李艳玲 葛凤培 林民 GAO Shang;LI Yan-ling;GE Feng-pei;LIN Min(College of Computer Science and Technology,Inner Mongolia Normal University,Hohhot 010022,China;Library,Beijing University of Posts and Telecommunications,Beijing 100876,China;Inner Mongolia Discipline Inspection and Supervision Big Data Laboratory,Hohhot 010015,China)

机构地区：[1]内蒙古师范大学计算机科学技术学院,内蒙古自治区呼和浩特010022 [2]北京邮电大学图书馆,北京100876 [3]内蒙古纪检监察大数据实验室,内蒙古自治区呼和浩特010015

出　　处：《计算机工程与设计》2023年第11期3447-3454,共8页Computer Engineering and Design

基　　金：国家自然科学基金项目(12204062、61806103、61562068);内蒙古自然科学基金项目(2022LHMS06001);内蒙古纪检监察大数据实验室开放课题基金项目(IMDBD2020013);内蒙古自治区“草原英才”工程青年创新创业人才基金项目(Q2017027);内蒙古自治区科技计划基金项目(JH20180175);内蒙古自治区高等学校科学技术研究基金项目(NJZY21578、NJZY21551);内蒙古师范大学校级基金项目(2022JBQN106、2022JBQN111)。

摘　　要：针对身份证信息识别中的文字定位和文字识别问题,提出基于改进卷积神经网络的身份证信息检测与识别方法。基于文字识别系统的主流框架并进行以下改进:针对目标检测网络定位文本精度低的问题,依据身份证的样式特点采用模板匹配方法进行文本定位。针对卷积核提取的特征表示能力弱的问题,采用残差模块结合卷积神经网络进行文字识别,使用预训练-微调范式解决证件数据集匮乏问题。针对文本间距离计算效率低的问题,采用集束搜索和语义处理优化识别结果。实验结果表明,该算法与对比方法相比,识别准确率有较大提升。For the text localization and text recognition problems in identity(ID)card information recognition,an ID card information detection and recognition method based on improved convolutional neural network was proposed.Based on the mainstream framework of text recognition system,the following improvements were made.For the problem of low accuracy of text localization through target detection network,a template matching method was used to locate the text based on the style characteristics of ID card.To address the problem of weak feature representation ability extracted by convolutional kernel,the residual module combined with convolutional neural network was used for text recognition,and the pre-trained fine-tuning paradigm was used to solve the problem of lack of ID dataset.To address the problem of low efficiency of inter-textual distance calculation,beam search and semantic processing were used to optimize the recognition results.Experimental results show that the algorithm has improvement in recognition accuracy compared with the comparison methods.

关 键 词：光学字符识别 文字识别 身份证识别 预训练 深度卷积神经网络 残差连接 模板匹配 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]