改进Faster-RCNN的场景文本检测算法

Improved Faster-RCNN for Scene text detection algorithm

由于文本信息比较复杂,字体大小不一,具有的像素信息较少,导致在特征提取阶段得到的特征图表达不充分,文本检测的准确率不高。针对以上问题,提出一种改进的Faster-RCNN文本检测方法。首先构建特征提取模块,用ResNet-101代替传统的VGG16网络提取图像特征;再融合特征金字塔的多尺度文本特征提取网络,在此基础上加入特征金字塔注意力模块;最后输入RPN层进行分类与边框回归。实验结果表明,改进后的Faster-RCNN比仅仅加入特征融合或特征金字塔注意力模块对文本检测效果提升明显。

The text information is more complex,the font size is different,the pixel information itself is less,the feature map fea⁃tures obtained in the feature extraction stage are not fully expressed,and the accuracy of detection is low.In view of the above prob⁃lems,this paper proposes an improved Faster-RCNN text detection method.Firstly,the feature extraction module is constructed,and the image features are extracted by replacing the traditional VGG16 network with ResNet-101;then the multi-scale text feature extrac⁃tion network of the feature pyramid is fused,and the feature pyramid attention module is added on the basis of the feature extraction net⁃work and feature fusion;and finally the RPN layer is input for classification and border regression.Experimental results show that the improved Faster-RCNN can significantly improve the text detection effect than the feature fusion or feature pyramid attention module.