BHS-CTPN:一种自然场景下的化验单文字检测方法  

BHS-CTPN:A TEXT DETECTION METHOD FOR LABORTATORY SHEETS IN NATURAL SCENES

作　　者：黄庆浩 吕学强[1] 何健 游新冬[1] Huang Qinghao;LV Xueqiang;He Jian;You Xindong(Beijing Key Laboratory of Internet Culture and Digital Dissemination Research,Beijing Information Science and Technology University,Beijing 100101,China;Beijing Luoqi Wisdom Medical Technology Co.,Ltd.,Beijing 100015,China;Tsinghua University Internet Industry Research Institute,Beijing 100084,China)

机构地区：[1]北京信息科技大学网络文化与数字传播北京市重点实验室,北京100101 [2]北京洛奇智慧医疗科技有限公司,北京100015 [3]清华大学互联网产业研究院,北京100084

出　　处：《计算机应用与软件》2022年第9期206-212,共7页Computer Applications and Software

基　　金：国家自然科学基金项目(61671070);国家语委重点项目(ZDI135-53);自然科学重点实验室基金(6142006190301);北京成像技术高精尖创新中心项目(BAICIT-2016003)。

摘　　要：针对存在的化验单文字检测方法效果无法达到实用标准以及敏感信息区域无法有效过滤的问题,提出一种自然场景下的化验单文字检测方法(BHS-CTPN)。BHS-CTPN方法引入BRISK、Hough、Sauvola等一系列预处理方法进行化验单矫正、敏感信息区域去除和图像增强。同时在特征提取时卷积核的设计、anchor设定和文本框合并等方面来改进CTPN网络模型。BHS-CTPN方法与CTPN模型相比,准确率、召回率、F1值分别提升8%、10%、9%;与目前化验单文字检测领域效果最好的华为API接口相比,准确率、召回率、F1值分别提升6%、3%、5%。大量实验结果表明,BHS-CTPN方法能够有效准确地检测到自然场景下化验单文本位置,为后期文字识别和化验单解读奠定坚实的基础。Aimed at the problems that the current existing laboratory sheets text detection methods cannot reach application standards and sensitive information areas can not be effectively filtered,a test sheet text detection method(BHS-CTPN)in natural scenes is proposed.BHS-CTPN method introduced a series of preprocessing methods such as BRISK,Hough,Sauvola to correct laboratory sheets,remove sensitive information regions and enhance images.The CTPN network model was improved in the design of convolution kernel,anchor setting and text box merging during feature extraction.Compared with CTPN model,BHS-CTPN method has increased its accuracy,recall rate and F1 value by 8%,10%and 9%respectively.Compared with Huawei API interface,which has the best effect in the field of for laboratory sheets text detection,BHS-CTPN method has increased its accuracy,recall rate and F1 value by 6%,3%and 5%respectively.Substantial experimental results show that BHS-CTPN method can effectively and accurately detect the text position of laboratory sheets in natural scenes,which lays a solid foundation for later text recognition and laboratory sheets interpretation.

关 键 词：自然场景 化验单 文字检测 预处理 CTPN 

分 类 号：TP3[自动化与计算机技术—计算机科学与技术]