[1]张庭瑞,方承志,徐国钦,等.基于多分支特征融合的自然场景文本检测算法[J].计算机技术与发展,2024,34(02):142-147.[doi:10. 3969 / j. issn. 1673-629X. 2024. 02. 021]
　ZHANG Ting-rui,FANG Cheng-zhi,XU Guo-qin,et al.Natural Scene Text Detection Algorithm Based on Multi-branch Feature Fusion[J].,2024,34(02):142-147.[doi:10. 3969 / j. issn. 1673-629X. 2024. 02. 021]

点击复制

基于多分支特征融合的自然场景文本检测算法()

分享到：

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

卷:

34

期数:

2024年02期

页码:

142-147

栏目:

人工智能

出版日期:

2024-02-10

文章信息/Info

Title:

Natural Scene Text Detection Algorithm Based on Multi-branch Feature Fusion

文章编号:

1673-629X(2024)02-0142-06

作者:

张庭瑞; 方承志; 徐国钦; 陈睿霖

南京邮电大学 电子与光学工程学院、柔性电子(未来技术)学院,江苏 南京 210023

Author(s):

ZHANG Ting-rui; FANG Cheng-zhi; XU Guo-qin; CHEN Rui-lin

School of Electronic and Optical Engineering,School of Flexible Electronics ( Future Technology) , Nanjing University of Posts and Telecommunications,Nanjing 210023,China

关键词:

文本检测; EAST 算法; 浅层特征增强; 循环十字交叉注意力; 损失函数

Keywords:

text detection; EAST; shallow feature enhancement; recurrent criss-cross attention; loss function

分类号:

TP391

DOI:

10. 3969 / j. issn. 1673-629X. 2024. 02. 021

摘要:

EAST 算法是一种高效而准确的场景文本检测算法,但是由于受到感受野的限制,导致在检测小文本时容易出现误检、漏检现象,在检测较长文本时缺乏一定的完整性。 针对以上问题,提出一种基于多分支特征融合的自然场景文本检测算法。 该算法以 EAST 算法为基础,引入并改进了浅层特征增强模块( RFB-s) ,在避免小文本信息损失的前提下,增大浅层网络的感受野改善浅层特征语义信息不足的问题,增强对小文本定位的准确性。 引入并改进了循环十字交叉注意力模块(RCCAM) ,使得特征图中的每个像素能够以非常有效的方式捕获全图像的上下文信息,提高对长文本的检测能力。同时针对回归任务,采用 Dice Loss 作为损失函数,解决正负样本占比不均衡问题。 采用 EIoU 来提高回归的效果,得到更为精准的文本框。 该算法在 ICDAR2015 和 MSRA-TD500 数据集上进行测试,均获得了不错的检测效果。 表明了该算法能够有效地对自然场景文本进行检测,提高了检测的准确率。

Abstract:

EAST is an efficient and accurate scene text detection algorithm,but due to the limitation of receptive field,it is prone to falsedetection and missed detection when detecting small text,?

and lacks certain integrity when detecting long text. Aiming at the aboveproblems,a natural scene text detection algorithm based on multi-branch feature fusion is proposed. Based on?

the EAST,the proposed algorithm introduces and improves the shallow feature enhancement module ( RFB - s) , and increases the receptive field of the shallownetwork to improve?

the problem of insufficient semantic information of shallow features on the premise of avoiding the loss of small textinformation to enhance the accuracy of small text positioning. The Recurrent Criss-Cross Attention Module ( RCCAM) is introduced andimproved,so that each pixel in the feature map can capture the contextual information of the full image in a very effective way,improvingthe detection ability of long text. At the same time,for the regression task,Dice Loss is used as the loss function to solve the problem ofunbalanced proportion of positive and negative samples. EIoU is used to improve the effect of regression and get a more accurate textbox. The proposed algorithm was tested on the ICDAR2015 and MSRA-TD500 datasets,and both achieved good detection results. It isshowed that the proposed algorithm can effectively detect natural scene text and improve the accuracy of detection.

常用功能