自然场景下的任意方向文本检测算法研究

Arbitrary direction text detection algorithm in natural scene

针对自然场景中的文本检测任务目前存在的几个难点展开重点研究,在已有文本检测算法的基础上,提出了一种基于锚的任意方向文本检测方法,对传统的区域建议网络(region proposal network,RPN)进行改进,使之可以生成含有方向角度信息的锚框。方法还利用仿射变换将RPN生成的带有旋转角度的候选区域映射到特征图上,并使用双线性插值来消除量化带来的误差。该方法多用于检测自然场景中带有一定倾斜角度的文本。实验结果表明,在ICDAR2015数据集上,本文提出的基于锚的任意方向文本检测方法的精确率为0.88,召回率为0.77,F-measure为0.82,优于大部分现有的基于锚的文本检测方法,并且该方法对小尺寸文本的检测效果极佳。

This paper focuses on several difficulties existing in natural scene text detection.Based on the existing text detection algorithms,an anchor-based arbitrary direction text detection method is proposed.The traditional region proposal network(RPN)is improved to generate anchor box with direction and angle information.In addition,the text proposals with rotation angles generated by RPN are projected to feature map using affine transformation,and the errors caused by quantization are eliminated using bilinear interpolation.The method proposed in this paper is mostly applied to detect texts with an oblique angle in natural scenes.The experimental results,on the ICDAR2015 dataset,show that the accuracy of the anchor-based arbitrary direction text detection method proposed in this paper is 0.88,the recall is 0.77,and the F-measure is 0.82,which is better than most existing anchor-based text detection methods,and this method has excellent effect on small-scale text detection.