Synthesizing Realistic Data for Table Recognition

arXiv:2404.11100v1 Announce Type: new
Abstract: To overcome the limitations and challenges of current automatic table data annotation methods and random table data synthesis approaches, we propose a novel method for synthesizing annotation data specifically designed for table recognition. This method utilizes the structure and content of existing complex tables, facilitating the efficient creation of tables that closely replicate the authentic styles found in the target domain. By leveraging the actual structure and content of tables from Chinese financial announcements, we …

abstract annotation arxiv challenges cs.cv cs.lg current data data annotation limitations novel random recognition synthesis table tables type