[1]邵慧敏,张太红*.基于 CTPN 神经网络对营业执照文字检测模型[J].计算机技术与发展,2021,31(01):94-97.[doi:10. 3969 / j. issn. 1673-629X. 2021. 01. 017]
　SHAO Hui-min,ZHANG Tai-hong*.Text Detection Model for Business License Based on CTPN Neural Network[J].,2021,31(01):94-97.[doi:10. 3969 / j. issn. 1673-629X. 2021. 01. 017]
点击复制

基于 CTPN 神经网络对营业执照文字检测模型()
分享到：

《计算机技术与发展》[ISSN:1006-6977/CN:61-1281/TN]

卷:
31
期数:
2021年01期
页码:
94-97
栏目:
图形与图像
出版日期:
2021-01-10

文章信息/Info

Title:
Text Detection Model for Business License Based on CTPN Neural Network
文章编号:
1673-629X(2021)01-0094-04
作者:
邵慧敏; 张太红*
新疆农业大学 计算机与信息工程学院,新疆 乌鲁木齐 830001
Author(s):
SHAO Hui-min; ZHANG Tai-hong*
School of Computer and Information Engineering,Xinjiang Agricultural University,Urumqi 830001,China
关键词:
营业执照; 文字检测; Tensorflow; Opencv; CTPN
Keywords:
business license; word detection; Tensorflow; Opencv; CTPN
分类号:
TP183
DOI:
10. 3969 / j. issn. 1673-629X. 2021. 01. 017
摘要:
对于复杂背景图片的文字识别,首先要做的就是定位目标文字的位置,即文字检测。 想要文字识别率高,那对文字检测的准确度的要求就非常高了。 传统的 RPN(region proposal network)神经网络在文字检测领域的研究已经很成熟,但 RPN 神经网络在营业执照水平文字检测的准确度上不是很理想。 而基于 CTPN(connectionist text proposal network)神经网络的文字检测模型明显提高了营业执照水平文字检测的正确率,但用于项目中的话,准确率还是远远不够的。 该文是以最新的营业执照作为研究对象,由于检测的图片易受光照和采集设备的影响,加上营业执照的背景比较复杂,所以能够准确地检测到目标文字的位置就非常具有挑战性。 文中是通过 CTPN 神经网络模型来检测出营业执照中水平文字所在的位置,用矩形框来标注,也就是横向水平检测。 目前开源的 CTPN 模型,都是基于某种数据集来训练的,所以对营业执照的文字检测效果就很差,因此该文使用 2 000 张营业执照图像作为实验数据,进行 10 000 迭代训练 CTPN 模型,最终能够准确地检测到营业执照中目标文字的位置,供项目使用。
Abstract:
For text recognition of complex background images,the first thing to do is to locate the location of the target text,that is,text detection. To let the text recognition rate is high,the accuracy of text detection requirements are quite high. The traditional RPN (region proposal network) neural network has been very mature in the field of text detection,but the accuracy of RPN neural network in text detection at the level of business license is not ideal. While the text detection model based on CTPN (connectionist text proposal network) neural network significantly improves the accuracy of text detection at the level of business license,but the accuracy is far from enough when applied to the project. We focus on the latest business license as the research object. Since the detected pictures are susceptible to? the influence of lighting and acquisition equipment,and the background of the business license is complex,it is highly challenging to accurately detect the location of the target text. The position of horizontal text in the business license is detected by the CTPN neural network model, which is marked by a rectang-ular box,that is,horizontal detection. The CTPN model of open source is based on some data sets to train,so the text detection effect of business license is poor. We will use the 2 000 business license image as the experi-mental data,the 10 000 iteration training CTPN model,to finally accurately detect the location of the target text in the business license for the use of the project.
更新日期/Last Update: 2020-01-10