文献类型：期刊文献

中文题名：基于旋转变量的任意方向文本检测算法

英文题名：Arbitrary-oriented text detection algorithm based on rotation variables

作者：张兴全[1];叶西宁[1]

机构：[1]华东理工大学信息科学与工程学院,上海200237

年份：2020

卷号：41

期号：5

起止页码：1253

中文期刊名：计算机工程与设计

外文期刊名：Computer Engineering and Design

收录：CSTPCD;;北大核心:【北大核心2017】；

基金：国家自然科学基金项目(60974066)。

语种：中文

中文关键词：文本检测;旋转变量;特征融合;边缘细化;小尺度文本

外文关键词：text detection;rotation variables;feature fusion;side refinement;small scale text

摘要：针对CTPN算法不能检测倾斜文本和小尺度文本检测效果差的问题,提出一种基于旋转变量的改进文本检测算法(CTPN-R)。通过增加旋转角度预测层,将文本行拆分成一系列带旋转角度的anchor用于网络训练解决文本倾斜问题,加入特征融合层构建特征图金字塔以检测多尺度文本,改进边缘细化层的输出变量和回归方法,优化文本连接。CTPN-R在倾斜文本数据集MSRA-TD500上的检测综合性能指标F-measure达到了77.5%,比CTPN提高了38%。仿真结果表明,CTPN-R对实际场景中的文本图像检测精度高、实时性好,可以检测任意大小、任意方向文本。
Aiming at the problems that CTPN algorithm can not detect tilted text and it shows poor effects on the detection of small scale text,an improved text detection algorithm based on rotation variable(CTPN-R)was proposed.By adding the rotation angle prediction layer,the text line was split into a sequence of anchors with rotation angle for network training to solve the tilted text problem.The feature fusion layer was added to construct the feature map pyramid to detect multi-scale text,the output variables and regression methods of the side-refinement layer were improved and the text connection was optimized.CTPN-R achieves 77.5%of F-measure on the oblique text dataset MSRA-TD500,which is 38%higher than CTPN.The simulation results show that CTPN-R can detect text images in real scene with high precision and real-time performance,and it can detect text of any size and direction.