Rethinking detection based table structure recognition for visually rich document images

Detection models have been extensively employed for the Table Structure Recognition (TSR) task, aiming to convert table images into structured formats by detecting table components such as Columns and Rows. However, prevailing detection-based TSR models usually cannot perform well regarding cell-level metrics, such as TEDS, and the reasons hindering their performance are not thoroughly explored. Therefore, we first examine the underlying reasons impeding these models’ performance and find that the key issues are the improper problem formulation, the mismatch issue of detection and TSR metrics, the inherent characteristics of detection models, and the influence of local and long-range feature extraction. Based on these findings, we propose a tailored Cascade R-CNN based solution by introducing a new problem formulation, tuning the proposal generation, and applying deformation convolution and the proposed Spatial Attention Module. The experimental results show that our proposed model can improve the base Cascade R-CNN model by 19.32%, 11.56%, and 14.77% on the SciTSR, FinTabNet, and PubTables1M datasets regarding the structure-only TEDS, achieving state-of-the-art performance, demonstrating that our findings can serve as a valuable guide for enhancing detection-based TSR models. Our code and pre-trained models are public available 1 1 https://github.com/uobinxiao/CascadeTSRDet . . • Comprehensive analysis of the reasons impeding existing detection-based Table Structure Recognition (TSR) solutions. • Proposing a simple and effective state-of-the-art detection-based TSR solution. • Providing a general design guidance for all detection-based TSR solutions.