改进EAST算法的游戏场景文本检测  

Game Scene Text Detection Based on Improved EAST Algorithm

作　　者：池凯 赵逢禹[1] CHI Kai;ZHAO Feng-yu(School of Optical-Electrical&Computer Engineering,University of Shanghai for Science&Technology,Shanghai 200093,China)

机构地区：[1]上海理工大学光电信息与计算机工程学院,上海200093

出　　处：《小型微型计算机系统》2020年第10期2189-2193,共5页Journal of Chinese Computer Systems

基　　金：国家自然科学基金项目(61202376,61003031)资助。

摘　　要：文本识别是模式识别应用的一个重要领域,其中文本检测是文本识别的关键所在.EAST是目前较为新颖的自然场景文本检测算法,能够高效准确地实现文本定位,但是仍存在长文本检测不佳,感受野不够大的问题.本文针对复杂游戏场景及传统算法定位问题,对EAST算法进行了改进优化,在EAST网络结构中加入BLSTM神经网络,并优化了shrink_poly权重策略以提高感受野,提高文本定位效果.实验结果表明,该方案能高效地对游戏场景文本进行检测,极大地提高了检测准确率.Text recognition is an important field of pattern recognition applications,in which text detdction is the key to text recognition.Though EAST is a relatively new natural scene text detection algorithm that can efficiently and accurately implement text positioning,there are still problems with poor long text detection and insufficient receptive field.This article focused on the complex game scenarios and traditional algorithm positioning problems,and the EAST algorithm is improved by addings a BLSTM neural network to the EAST network structure,and the shrink_poly weighting strategy to improve the receptive field and improve the text positioning effect.The experimental results show that this scheme can efficiently detect the text of the game scene and greatly improve the detection accuracy.

关 键 词：游戏场景 文本检测 EAST 感受野 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]