基于边缘特征增强的任意形状文本检测网络

A New Arbitrary-shaped Text Detection Network by Reinforcing Edge Features

在场景文本检测方法中,文本实例的边缘特征与其他特征在大多数模型中都是以同样的方式进行处理,而准确检测相邻文本边缘区域是正确识别任意形状文本区域的关键之一.如果对边缘特征进行增强并使用独立分支进行建模,必能有效提高模型的标识准确率.为此,提出了三个用以增强边缘特征的网络模块.其中,浅层特征增强模块可有效增强包含更多边缘特征的浅层特征;边缘区域检测分支将普通特征和边缘特征进行区分以对目标的边缘特征进行显式建模;而分支特征融合模块可将两种特征在识别过程进行更好的融合.在将这三个模块引入渐进尺度扩张网络(Progressive scale expansion network, PSENet)之后,相关消融实验表明这三个模块的单独使用及其组合均可进一步增加网络的预测准确率.此外,在三个常用公开数据集上与其他十个最新模型的比较结果表明,改进后得到边缘特征增强网络(Edge-oriented feature reinforcing network, EFRNet)的识别结果具有较高的F1值.

In the detection of scene texts areas,the text instances'edge features are processed in the same way as other features.Nevertheless,the accurate detection of adjacent text edges is crucial in the correct identification of arbitrary-shaped text regions in natural scenes.Obviously,the identification accuracy increases if edge features can be enhanced and modeled through independent branches in the network.To this end,three network modules are proposed to enhance the edge features in this paper.These modules are the shallow feature enhancement module which effectively enhances the shallow features with more edge features,the edge region detection module which decouples the original features into edge features and text features to explicitly model the edge features of the object,and the branch feature fusion module which effectively fuses these two types of features in the recognition process.After the proposed modules are added to the progressive scale expansion network(PSENet),the ablation experiments show that both the independent application and the synthetic application of these modules increase the prediction accuracy.In addition,the comparison experiments on three commonly used public datasets with ten state-of-the-art methods show that the improved edge-oriented feature reinforcing network(EFRNet)has higher F1-measure accuracy.