基于BLSTM网络的改进EAST文本检测算法  

Improved EAST Natural Scene Text Location Algorithm Based on BLSTM Network

作　　者：郭闯 邱晓晖[1] GUO Chuang;QIU Xiao-hui(School of Communication and Information Engineering,Nanjing University of Posts and Telecommunications,Nanjing 210003,China)

机构地区：[1]南京邮电大学通信与信息工程学院,江苏南京210003

出　　处：《计算机技术与发展》2020年第7期21-24,共4页Computer Technology and Development

基　　金：江苏省自然科学基金(BK2011789)。

摘　　要：尽管前人在文本检测和文本识别方面已经取得了显著的研究进展,但是在场景文本检测方面仍然存在着较大的不足。即使是深度学习模型,也不会达到很好的性能。因为整体性能取决于流程中的多个阶段和组件的相互作用。基于深度学习神经网络模型的EAST算法可以在进行场景文本检测时避免传统文本检测方法不必要的中间步骤(例如候选区域和字分区域),从而得到了快速准确的检测效果,准确率和召回率都有大幅度的提高。然而由于其感受野范围较短,对长文本的检测效果仍存在问题,因此文中对EAST算法进行改进,在EAST算法的基础上,引入BLSTM网络,提高其感受野,增强文本定位的效果。实验结果表明,该算法在ICDAR2015文本定位任务的召回率为78.07%,准确率为85.10%,F-score为81.64%,优于经典EAST算法。Although the predecessors have made remarkable research progress in text detection and text recognition,there are still some deficiencies in scene text detection. Even the deep learning model will not achieve ideal performance. Because the overall performance depends on the interaction of multiple stages and components in the process. EAST algorithm based on deep learning neural network model can avoid unnecessary intermediate steps in traditional text detection methods(such as candidate regions and word sub-regions) when performing scene text detection,so as to obtain rapid and accurate detection effect,with a significant improvement in accuracy and recall rate. However,due to the short range of receptive fields,there are still problems in the detection of long text. Therefore,we improve the EAST algorithm,based on which BLSTM network is introduced to improve its sensing field and enhance the effect of text location. The experiment shows that the recall rate,accuracy and F-score of the proposed algorithm are 78.07%,85.10% and 81.64% respectively,which are better than the classical EAST algorithm.

关 键 词：文本定位 EAST BLSTM 感受野 自然场景 

分 类 号：TP301.6[自动化与计算机技术—计算机系统结构]