表格结构识别网络FcaTGRNet

Table structure recognition network FcaTGRNet

如今重要的数据大都通过表格直接展示,在分析数据时,除数据和文字信息外,表格结构信息也很重要,要想更加准确、便捷地分析表格数据,能否自动准确地识别表格结构至关重要。现在有很多识别表格结构的方法,对表格单元格检测的准确率也都比较理想,但是很少有方法在特征提取方面做出改进。为了证明图像特征提取在表格结构识别过程中的重要性,针对这一问题在TGRNet的基础上引入了多频谱注意力机制,以便更好地提取图像多个频谱上的特征,使得到的特征更加全面。在公开数据集上的实验结果显示,所做改进较原方法在表格结构识别的单元格空间位置和逻辑位置检测的准确率有所提升。

Nowadays, most important data are directly displayed through tables. In the analysis of data, in addition to data and text information, table structure information is also very important. In order to analyze table data more accurately and conveniently, whether the table structure can be automatically and accurately identified is very important. At present, there are many methods to recognize table structure, and the accuracy of table cell detection is ideal, but few methods make improvements in feature extraction. In order to prove the importance of image feature extraction in the process of table structure recognition, this paper introduces a Frequency channel attention Networks(FcaNet)based on TGRNet to solve this problem, so as to better extract the features in multiple spectrum of the image and make the features more comprehensive. Experimental results on open datasets show that the proposed method has improved the accuracy of spatial location and logical location detection in table structure recognition compared with the original method.