基于NLA-EAST的自然场景文本检测方法  

Natural Scene Text Detection Based on NLA-EAST

作　　者：姚焙继 朱玉全[1] 岑燕妮 YAO Beiji;ZHU Yuquan;CEN Yanni(School of Computer Science and Communication Engineering,Jiangsu University,Zhenjiang 212013)

机构地区：[1]江苏大学计算机科学与通信工程学院,镇江212013

出　　处：《计算机与数字工程》2021年第7期1420-1425,共6页Computer & Digital Engineering

摘　　要：近年来,基于深度学习的场景文本检测算法层出不穷,对于EAST在自然场景中对长文本和较大文本检测不准确,存在容易出现误检漏检的问题。论文提出一种基于NLA-EAST网络(Non-Local Attention-An Efficient and Accurate Scene Text Detector)上的新颖的文本检测算法,通过ASPP空洞卷积来扩大感受野,来获得更大感受野的上下文信息。并且通过结合EAST和非局部注意力机制来精确定位文本边界,准确检测自然场景下的文本位置,克服了EAST对于较大文本和长文本的漏检和误检。对提出的方法进行了数据集测试,在文本定位精度方面由于竞争方法,在ICDAR 2015数据集中,F值达到了84.5%,在天池数据集上,F值达到了84.82%。In recent years,scene text detection algorithms based on deep learning have emerged in an endless stream.For EAST in the natural scene,the detection of long text and large text is inaccurate,and there is a problem that misdetection is likely to occur.This paper proposes a novel text detection algorithm based on NLA-EAST(Non-Local Attention-An Efficient and Accu⁃rate Scene Text Detector),which expands the receptive field through ASPP cavity convolution to obtain the contextual information of the larger greater receptive field.And by combining EAST and non-local attention mechanism to accurately locate text boundar⁃ies,this method can accurately detect the position of text in natural scenes,and overcomes EAST's missed detection and false detec⁃tion of large text and long text.The data set is tested on the proposed method.In terms of text localization accuracy,the F value reaches 84.5%in the ICDAR 2015 data set,and the F value reaches 84.82%in the Tianchi data set.

关 键 词：场景文本检测 长文本 较大文本 NLA-EAST 非局部注意力 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]