[1]闫 炜,周凌柯.基于 CTPN 算法检测试卷手写成绩数字的方法[J].计算机技术与发展,2022,32(S1):61-65.[doi:10. 3969 / j. issn. 1673-629X. 2022. S1. 014]
　YAN Wei,ZHOU Ling-ke.A Method of Detecting Handwritten Score Numbers in Test Papers Based on CTPN Algorithm[J].,2022,32(S1):61-65.[doi:10. 3969 / j. issn. 1673-629X. 2022. S1. 014]
点击复制

基于 CTPN 算法检测试卷手写成绩数字的方法()
分享到：

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

卷:
32
期数:
2022年S1期
页码:
61-65
栏目:
应用前沿与综合
出版日期:
2022-12-11

文章信息/Info

Title:
A Method of Detecting Handwritten Score Numbers in Test Papers Based on CTPN Algorithm
文章编号:
1673-629X(2022)S1-006105
作者:
闫 炜; 周凌柯
南京理工大学 自动化学院,江苏 南京 210000
Author(s):
YAN Wei; ZHOU Ling-ke
School of Automation,Nanjing University of Technology,Nanjing 210000,China
关键词:
智能录分系统; 成绩检测; 图像预处理; CTPN; ResNet
Keywords:
intelligent recording system; word detection; image preprocessing; CTPN; ResNet
分类号:
TP301
DOI:
10. 3969 / j. issn. 1673-629X. 2022. S1. 014
摘要:
试卷批阅的手写成绩文本自动检测是智能考试系统中的一个重要环节,试卷手写成绩识别的首要步骤是目标数字的定位,即手写成绩自动检测。 据调查,试卷批阅的手写成绩存在差异化现象且试卷答题纸背景复杂,影响了自动检测的准确率。 传统的 CTPN( connection text proposal network) 对于水平文本检测有着不错的效果,但直接应用于试卷手写成绩检测,准确率仍达不到要求。 为了提高手写文本检测的准确率,针对检测过程中答题纸内容、照片拍摄背景复杂等问题,提出了一种基于 CTPN 的改进算法,用来检测手写批阅成绩。 使用颜色提取、中值滤波、膨胀等方法对输入的试卷图片进行预处理,提取批阅成绩内容;然后使用深层卷积网络代替 VGG16 作为特征提取网络,提升特征提取的能力,检测出成绩栏所在位置;最后使用一种新的成绩栏分割策略,把试卷中每一项的总得分独立提取出来。 经过实验发现,提出的预处理和改进的 CTPN 算法可以更为准确地检测到试卷上的成绩区域,在精确度、召回率方面比传统 VGG16 方法有着明显改善。
Abstract:
Automatic text detection of examination paper marking is an important part of intelligent examination system. For examinationpaper recognition,the first thing to do is to locate the target number. According to the survey, the handwritten text differentiationphenomenon of examination paper marking is serious,and the background of examination paper is complex,which affects the accuracy ofautomatic detection. The traditional CTPN ( connection text proposal network) has a good effect for horizontal text detection,but in theapplication and test paper performance detection,the accuracy still cannot meet the requirements. In order to improve the accuracy ofhandwritten text detection,we propose an improved algorithm based on CTPN to detect the handwritten marking results. First,we usecolor extraction,median filtering, expansion and other methods to preprocess the input test paper pictures, extract the content of thegrading results,and then use deep convolution network instead of VGG16 as the feature extraction network to improve the ability offeature extraction and detect the position of the score bar. Finally,a new strategy of score column segmentation is used to extract the totalscore of each item independently. The experimental results show that the pre-processing and improved CTPN algorithm proposed in thispaper can accurately detect the score area on the test paper,and the accuracy and recall rate are significantly improved compared with thetraditional VGG16 method.
更新日期/Last Update: 2022-06-10