基于改进EAST算法的专利附图标记检测  

Patent label detection based on improved EAST algorithm

作　　者：范楠 肖诗斌[1,2] 王焕鹏 施水才 FAN Nan;XIAO Shibin;WANG Huanpeng;SHI Shuicai(TRS Sof tware Opening Laboratory,Beijing Information Science&Technology University,Beijing 100101,China;Beijing TRS Information Technology Co.,Ltd.,Beijing 100101,China)

机构地区：[1]北京信息科技大学TRS软件开放实验室,北京100101 [2]拓尔思信息技术股份有限公司,北京100101

出　　处：《北京信息科技大学学报（自然科学版）》2021年第4期95-100,共6页Journal of Beijing Information Science and Technology University

摘　　要：为充分利用发明专利和实用新型专利的附图信息,进一步研究利用专利附图提高专利检索的效率,提出一种基于改进EAST算法的专利附图标记检测方法。对专利附图标记进行检测时,改进特征提取阶段的骨干网络为ResNet50,并且融合空间注意力和通道注意力机制,经过多维度附图标记特征提取和融合,实现对专利附图标记的检测。与原EAST算法相比,改进的算法在专利附图标记检测时的精确率要高1.12%,召回率高4.7%,F1值高3%。实验表明,改进的EAST算法能够有效准确地检测专利附图标记的位置,并且附图标记检测的召回率有显著提升。In order to make full use of the drawing information of invention patents and utility model patents,and improve the efficiency of patent retrieval by the use of patent drawing,a patent part label detection method based on the improved EAST algorithm was proposed.For the detection of patent part label,the backbone network of the improved feature extraction stage was ResNet50,the spatial attention and channel attention mechanisms were fused,and the detection of patent part label was realized after multi-dimensional part label feature extraction and fusion.Compared with the original EAST algorithm,the improved algorithm has 1.12%higher accuracy,4.7%higher recall,and 3%higher F1 score in patent part label detection.The experimental results show that the improved EAST algorithm can effectively and accurately detect the location of patent part label,and the recall rate of part label detection is significantly improved.

关 键 词：专利附图标记 文本检测 注意力机制 EAST 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]