基于分支ResNet-50的EAST场景文字检测算法  

Text detection algorithm in East scene based on branch Resnet-50

作　　者：龙言 施水才[1,2] 肖诗斌 LONG Yan;SHI Shuicai;XIAO Shibin(Computer School,Beijing Information Science&Technology University,Beijing 100192,China;TRS Information Technology Co.,Ltd,Beijing 100101,China)

机构地区：[1]北京信息科技大学计算机学院,北京100192 [2]拓尔思信息技术股份有限公司,北京100101

出　　处：《北京信息科技大学学报（自然科学版）》2020年第3期94-98,共5页Journal of Beijing Information Science and Technology University

摘　　要：针对自然场景文字检测任务中多尺度文本提取效果不佳的问题,改进EAST算法的ResNet基础网络单元,采用分支结构的特征提取方法,使得算法能够感知到更加深层次的文本特征。实验证明,在ICDAR2015数据集上,分支结构的EAST算法相比原始EAST算法,提高了1.18%的准确率。该算法对于较为复杂的文本检测任务表现出良好的检测性能。Aiming at the problem of poor multi-scale text extraction effect in the text detection in natural scene,the ResNet basic network unit of East algorithm is improved,and the feature extraction method of branch structure is adopted,so that the algorithm can perceive the in-depth text features.Experimental results show that,compared with the original EAST algorithm,the EAST algorithm of branch structure improves the accuracy of 1.18%in ICDAR2015 data set.The algorithm performs well in more complex text detection tasks.

关 键 词：自然场景 文字检测 分支结构 EAST 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]