Enhancing Table Recognition Model by Synthetic Image Generation

Recently, table recognition has been gaining significant research attention. In table recognition tasks, creating large-scale table datasets for training and evaluating remains one of the greatest obstacles. In this paper, we propose a novel data generation method that combines an analysis of table characteristics with Generative-Adversarial-Networks (GANs) to generate synthetic table datasets, including both table images and their corresponding ground truths. The results show that our method achieves a remarkable amount of time for two tasks generating images and automatic labeling, with approximately 0.03 seconds and 0.017 seconds for a table image respectively. Moreover, to prove the efficiency of generating labeled table images when the number of analyzed tables increases, the performance of table detection and table structure recognition is evaluated by fine-tuning the Detection Transformer (DETR) model.