基于DBNET与CRNN-CTC的自然环境文字识别系统  

NATURAL ENVIRONMENT CHARACTER RECOGNITION SYSTEMBASED ON DBNET AND CRNN-CTC

作　　者：郭浩 宁初明 韩寿松 李华莹 Guo Hao;Ning Chuming;Han Shousong;Li Huaying(66407 Unit PLA,Beijing 100089,China;System Engineering Research Institute,Military Academy of Sciences,Beijing 100020,China;Department of Vehicle Engineering,Army Armored Force Academys,Beijing 100071,China)

机构地区：[1]六六四〇七部队,北京100089 [2]军事科学院系统工程研究院,北京100020 [3]陆军装甲兵学院车辆工程系,北京100071

出　　处：《计算机应用与软件》2023年第9期132-136,共5页Computer Applications and Software

基　　金：国家自然科学基金项目(51875575)。

摘　　要：随着军队信息化建设的逐步深入,当前存在大量的文稿、单据、表格迫切需求高精度的文字识别处理。一般光学字符识别(OCR)技术如谷歌开源Tesseract引擎等对环境光源、字体,视角要求苛刻,仅能兼容扫描文档,难以实现自然环境下拍照使用需求,使用困难且准确率低。该文提出一种基于可微二值化网络(Differentiable Binarization Net, DBNET)进行场景文本分割,并采用CRNN-CTC进行文字识别的复合二阶段识别方法,可在复杂光源、多角度视角的自然场景中实现长序列多文字识别。该方法的识别效率及准确率远远高于传统OCR方法、对比其他深度学习方法亦存在较大优势。在该方法基础上进一步构建基于异步消息队列的并行架构服务器系统,使得技术得以进一步实用化。With the gradual deepening of military information construction,there are a large number of manuscripts,documents and forms,which urgently need high-precision character recognition processing.General optical character recognition(OCR)technologies,such as Google's open source Tesseract engine,have strict requirements on environmental light source,font and perspective,which can only be compatible with scanners,so it is difficult to meet the needs of photographing in natural environment,and it is difficult to use and has low accuracy.A differential binarization net(DBNET)is proposed for scene text segmentation,and CRNN-CTC for character recognition.Two-stage recognition method could be used in the complex light source,multi angle natural scene to achieve long sequence multi character recognition,the recognition efficiency and accuracy of this method was far higher than the traditional OCR method.It also has greater advantages compared with other deep learning methods.At the same time,this paper made a further study on the basis of this method,and constructed a parallel architecture server system based on asynchronous message queue,which made the technology more practical.

关 键 词：光学字符识别 深度学习 软件系统 

分 类 号：TP3[自动化与计算机技术—计算机科学与技术]