Scene Text Detection with Improved Region Proposal Network

Abstract

Scene text images have very complex and changeable features.Using region proposal network(RPN) to extract text rectangle position candidate boxes is an indispensable step,which can greatly improve the accuracy of text detection.However,recent studies show that the methods of regressing the center point,width and height of the text rectangular candidate boxes by minimizing the smooth L1 loss function would easily cause problems such as missing boundary information and inaccurate regression.Therefore,this paper proposes a scene text detection model based on improved region proposal network.First,the backbone network composed of the residual network and the feature pyramid network is used to generate a shared feature map.Then,an improved regression method and vertex-based loss function(Vertex-IOU) are used to generate a series of text rectangular candidate boxes on the shared feature map.Finally,ROI Align is used to convert these candidate boxes into fixed-size feature maps for bounding box regression in the fully connected layer.Through comparative experiments on ICDAR2015 dataset,the results show that the test effect is improved compared with other models,which proves the effectiveness of our model.

Keywords

• keywords deep learning|scene text detection|region proposal network|regression method|loss function