嵌入重评分机制的自然场景文本检测方法  

Text Detection in Natural Scenes Based on Embedded Re-Score Mechanism

作　　者：刘艳丽 王毅宏[2] 张恒 程晶晶 LIU Yan-li;WANG Yi-hong;ZHANG Heng;CHENG Jing-jing(School of Electronic Information,Shanghai Dianji University,Shanghai 201306,China;School of Information Engineering,East China Jiaotong University,Nanchang Jiangxi 330000,China)

机构地区：[1]上海电机学院电子信息学院,上海201306 [2]华东交通大学信息工程学院,江西南昌330000

出　　处：《计算机仿真》2023年第2期228-235,302,共9页Computer Simulation

基　　金：国家自然科学基金资助项目(61963017);上海市科技计划项目(23010501000);上海市教育科学研究项目(C2022056);教育部人文社会科学研究项目(22YJAZH145)。

摘　　要：针对自然场景文本检测中存在大量假阳性问题,提出了嵌入重评分机制的自然场景文本检测方法。引入实例分割网络(Mask R-CNN)作为基本框架,实现对自然场景中多方向、不规则文本的检测;设计文本掩膜重评分机制,通过预测文本掩膜的质量,将文本的语义类别信息与其对应的掩膜完整性信息相结合,重新评估文本掩膜的质量,精确了文本的候选区域;重新设计损失函数的作用范围。上述模型基于端到端训练,在ICDAR2013、ICDAR2015和Total-Text等数据集进行性能测试,结果表明,提出的方法有效的提高了字符分割的完整性,较之现有方法明显地提高了文本检测的准确率和召回率,更适合自然场景中的不规则文本的识别。In order to solve the problem of false positives in text detection of natural scenes,a text detection method based on embedded res-coring is proposed.This paper introduced the instance segmentation network(Mask RCNN)as the basic framework to realize multi-directional and irregular text detection in natural scenes.By predicting the quality of the text mask,and combining the semantic category information of the text with its corresponding mask integrity information,the quality of the text mask was re-evaluated,and the candidate regions of the text were precisely refined.The scope of the loss function was redesigned.The model was based on end-to-end training,and its performance was tested on ICDAR2013,ICDAR2015 and Total-Text datasets.The results show that this method effectively improves the integrity of character segmentation,significantly improves the accuracy and recall rate of text detection compared with existing methods,and is more suitable for irregular text recognition in natural scenes.

关 键 词：文本检测 文本识别 自然场景 实例分割 

分 类 号：TP399[自动化与计算机技术—计算机应用技术]