挖掘文本框位置特性的anchor-free自然场景文本检测  

Mining text box location feature for anchor-free text detection in natural scene images

作　　者：卢利琼 吴东 吴涛 刘瑶 Lu Liqiong;Wu Dong;Wu Tao;Liu Yao(School of Information Engineering,Lingnan Normal University,Zhanjiang Guangdong 524048,China;Guangdong Provincial Key Laboratory of Development&Education for Special Needs Children,Lingnan Normal University,Zhanjiang Guangdong 524048,China;Teaching Development&Quality Evaluation Center,Lingnan Normal University,Zhanjiang Guangdong 524048,China)

机构地区：[1]岭南师范学院信息工程学院,广东湛江524048 [2]岭南师范学院广东省特殊儿童发展与教育重点实验室,广东湛江524048 [3]岭南师范学院教学发展与质量测评中心,广东湛江524048

出　　处：《计算机应用研究》2021年第8期2556-2560,共5页Application Research of Computers

基　　金：国家自然科学基金资助项目(62005109);广东省自然科学基金资助项目(2017A030307030);岭南师范学院自然科学人才专项资助项目(ZL2021015);广东省特殊儿童发展与教育重点实验室资助项目(TJ202011)。

摘　　要：针对现有优秀的anchor-free文本检测方法只挖掘了文本框几何特性而没有考虑文本框位置特性且缺乏有效的过滤机制,提出了挖掘文本框位置特性的anchor-free自然场景文本检测方法。该方法以ResNet50作为卷积神经网络的主干网络,将多个不同尺寸的特征层融合后预测文本框的几何特性和位置特性,最后辅之以二层过滤机制得到最终的检测文本框。在公开的数据集ICDAR2013和ICDAR2011上F值分别达到了0.870和0.861,证明了该方法的有效性。The exiting excellent text detection methods only mine geometric feature of text boxes and lack an effective filtering mechanism.Aimed at this problem,this paper proposed an anchor-free text detection method through mining text box location feature.The method used ResNet50 as the backbone of CNN(convolution neural network),and then fused the features with different size to predict the geometric feature and location feature of text boxes.Finally,this method used a two filtering mechanism to obtain the final detection text boxes.The method got the value of F at 0.870 and 0.861 on ICDAR2013 and ICDAR2011 datasets,which demonstrate the effectiveness of mining text box location feature for anchor-free text detection in natural scene images.

关 键 词：自然场景图像 文本检测 位置特性 anchor-free 卷积神经网络 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]