| The knowledge base of the financial industry can solve the problem that the financial staffs spend more time when retrieving data.The tables in financial documents contain a large number of valuable financial attributes information,but financial tables often have a large number of complicated merged cells or lack of a large number of border lines.These problems make it difficult to accurately extract the structural information of financial tables.The current extraction of these information is called table structure recognition.In the table structure recognition,in order to detect and extract the structure of the table in the document,researchers often use heuristic rules or technology in the field of computer vision.With the development of deep learning,computer vision technology has become the mainstream.In the detection of table,in the task,the target detection method is mainly used to detect the coordinate range of the table in the document;in the table structure recognition task,the semantic segmentation method is mainly used to divide the row and column as the granularity,and then the row and column segmentation area.The overlapping part is analyzed to obtain the structural information and text content information of each cell.These solutions often do not work well on complex financial tables.In view of the lack of training data and the information extraction of financial charts,a data generation technology based on weakly supervised learning and a financial table structure recognition method based on instance segmentation are proposed.Its core idea is to change the minimum granularity in table recognition.The recognition of the grid area avoids the problem of merging of multiple rows and columns of data due to too close distance,and then restores the structural information of the entire table according to the recognition result.For the first time,the method of instance segmentation was used to identify the structure of financial table.The difficulties in identifying the structure of financial tables were analyzed,and the method of identifying financial forms by instance segmentation was proposed and implemented.In the data generation part,weakly supervised learning will be used to solve the problem of insufficient data in this field,avoiding a lot of manual labeling work.The design experiment tested and analyzed some of the complex tables in the 92 artificially marked real financial tables and the ICDAR2013 table dataset,and achieved a recognition precision rate of 0.9 on the manual marked financial dataset,which verified the effectiveness of the instance segmentation technology for the identification of financial tables. |
