基于DBNet改进的检务场景文本检测算法研究
Research on improved text detection algorithm for prosecutorial scenarios based on DBNet
于晓 1林世基1
作者信息
- 1. 天津理工大学电气工程与自动化学院,天津 300384
- 折叠
摘要
针对检务场景文本检测中,现有的检测算法仍存在误检率和漏检率高等问题.通过改进现有的特征提取网络,引入高效通道注意力和空间注意力模块CBAM,同时改进可微二值化函数,并将改进后的网络应用到检务场景文本检测当中.改进后的算法在ICDAR 2015数据集上的准确率、召回率及F值相较于改进前分别提升了2.2、5.4及4.2个百分点,达到了89.2%和63.6%及74.3%.实验数据表明,改进DBNet文本检测算法在收敛速度和检测精度上都有明显的提升.
Abstract
In the text detection of inspection scene,the existing detection algorithms still have the problems of high false detec-tion rate and high missing detection rate.By improving the existing feature extraction network,introducing the efficient channel at-tention and spatial attention module CBAM,and improving the differentiable binary function,the improved network is applied to the text detection of inspection scene.The accuracy,recall and F value of the improved algorithm on ICDAR 2015 data set in-creased by 2.2,5.4 and 4.2 percentage,respectively,to 89.2%,63.6%and 74.3%compared with those before the improvement.Ex-perimental data show that the improved DBNet text detection algorithm has a significant improvement in convergence speed and de-tection accuracy.
关键词
文本检测/检务场景/可微二值化/深度学习/CBAMKey words
text detection/prosecutorial scenes/sifferentiable binarization/deep learning/CBAM引用本文复制引用
基金项目
国家自然科学基金(61502340)
天津市自然科学基金(18JCQNJC01000)
出版年
2024
NETL
NSTL
万方数据