基于注意力机制的多方向文本检测  

Multi-directional text detection based on attention mechanism

作　　者：徐健[1] 郭湛澎 刘秀平[1] 陈博 闫焕营 XU Jian;GUO Zhanpeng;LIU Xiuping;CHEN Bo;YAN Huanying(School of Electronics and Information,Xi'an Polytechnic University,Xi'an,Shaanxi 710048,China;Municipal Robotel Robot Technology Co.,LTD,Shenzhen,Guangdong 518109,China)

机构地区：[1]西安工程大学电子信息学院,陕西西安710048 [2]深圳罗博泰尔机器人技术有限公司,广东深圳518109

出　　处：《光电子．激光》2023年第2期166-173,共8页Journal of Optoelectronics·Laser

基　　金：陕西省科技厅项目(2018GY-173);西安市科技局项目(GXYD7.5)资助项目。

摘　　要：针对多方向排列的文本因其尺度变化大、复杂背景干扰而导致检测效果仍不甚理想的问题,本文提出了一种基于注意力机制的多方向文本检测方法。首先,考虑到自然场景下干扰信息多,构建文本特征提取网络(text feature information ResNet50,TF-ResNet),对图像中的文本特征信息进行提取;其次,在特征融合模型中加入文本注意模块(text attention module, TAM),抑制无关信息的同时突出显示文本信息,以增强文本特征之间的潜在联系;最后,采用渐进扩展模块,逐步融合扩展前部分得到的多个不同尺度的分割结果,以获得精确检测结果。本文方法在数据集CTW1500、ICDAR2015上进行实验验证和分析,其F值分别达到80.4%和83.0%,比次优方法分别提升了2.0%和2.4%,表明该方法在多方向文本检测上与其他方法相比具备一定的竞争力。Aiming at the problem that the detection effect of multi-directional arrangement text is still not ideal due to its large scale change and complex background interference, this paper proposes a multi-directional text detection method based on attention mechanism. Firstly, considering that there is a lot of interference information in natural scenes, a text feature extraction network is constructed to extract the text feature information in the image;Secondly, a text attention module(TAM) is added to the feature fusion model to suppress irrelevant information while highlighting textual information to enhance potential connections between text features;Finally, a progressive expansion module is used to gradually fuse the segmentation results obtained from the pre-expansion part at several different scales to obtain accurate detection results.The method is experimentally validated and analysed on datasets CTW1500 and ICDAR2015,and its F-values reach 80.4% and 83.0% respectively, which are 2.0% and 2.4% better than the next best method, indicating that the method is competitive with other methods in multi-directional text detection.

关 键 词：场景文本检测 注意力机制 文本特征提取网络(TF-ResNet) 文本注意模块 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]