一种基于关系模型的场景文本检测方法  

A Method for Scene Text Detection Based on Relation Model

作　　者：陈旸 赵尔敦[1] 吴靖 CHEN Yang;ZHAO Erdun;WU Jing(School of Computer,Central China Normal University,Wuhan 430079)

机构地区：[1]华中师范大学计算机学院,武汉430079

出　　处：《计算机与数字工程》2023年第4期871-876,共6页Computer & Digital Engineering

摘　　要：基于深度学习的文本检测方法在自然场景文本检测中取得了令人瞩目的成效,但是目前的文本检测模型大多忽略了文本区域之间的关系特征。因此论文在深度学习的基础上,结合注意力机制中的关系模型,提出了一种基于关系模型的自然场景文本检测方法,该方法先利用自动设置锚的卷积神经网络来提取文本候选区域,然后利用关系模型结合候选文字区域之间的关系,从而准确提取文字区域。在ICDAR2013和ICDAR2015数据集上的实验结果表明,与其他算法相比,论文提出的算法能取得更鲁棒的综合性能,在自然场景文本上有较好的应用前景。At present,the text detection method based on deep learning has made many noteworthy achievements in the scene text detection.However,the relations between text regions are usually overlooked by most of the existed methods.Therefore,on the basis of deep learning,combined with the relational model of attention mecha-nism,a text detection method of natural scene based on relational model in this paper is proposed.The proposed method firstly gets the candidate text regions by the deep neural network which can learn to generate the anchor by itself,and then predicts the text regions by the relations among the candidate text regions with the help of the relation model.The experimental results on the ICDAR2013 and ICDAR2015 datasets show that compared with other algorithms,the algorithm proposed in this paper can achieve more robust comprehensive performance and has a better application prospect in natural scene text.

关 键 词：深度学习 关系模型 锚自生成网络 场景文本检测 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]