AGOE-TSR: adaptive gradient orientation extractor with balanced difficulty-aware loss for table structure recognition

Abstract

Table structure recognition (TSR) remains challenging due to geometric deformation, blurred boundaries, and multiscale cell variations. To address these issues, we propose AGOE-TSR, an end-to-end network using a parallel regression framework for robust table parsing. An Adaptive Gradient Orientation Extractor (AGOE) is introduced to enhance feature representation for curved and blurred cell boundaries by dynamically capturing gradient direction features. A Row-Column Separation (RCS) module incorporates global grid priors to enforce row and column contexts, thereby improving logical index prediction consistency. Furthermore, a Balanced Difficulty-aware Loss (BDL) adaptively reweights gradients based on cell area to alleviate multi-scale imbalance. Experiments demonstrate that AGOE-TSR achieves state-of-the-art Logical Accuracy on ICDAR-2013, TableGraph-24k, and WTW, and outperforms sequence-based methods on PubTabNet in TEDS-S, highlighting its robustness and generalization across diverse table datasets.