Multilingual Row Detection in Tables: Beyond TATR with YOLO, Faster R-CNN, and TEDS-S

Multilingual table extraction presents significant challenges due to structural variability, noisy datasets, and script-specific complexities. Traditional heuristic-based approaches, such as the Table Analysis and Recognition Tool (TATR), struggle to generalize across diverse scripts and irregular layouts. This research investigates deep learning–based alternatives for table row detection using the multilingual MUSTARD dataset. YOLO and Faster R-CNN models are evaluated in terms of accuracy, inference speed, and robustness across multilingual scripts. To enable holistic assessment, TEDS-S is employed to jointly evaluate structural alignment and content fidelity. Experimental results demonstrate that YOLO offers superior real-time performance, while Faster R-CNN achieves higher precision. Hybrid detection strategies further enhance row detection accuracy, highlighting the effectiveness of combining object detection models with structure-and-content-aware evaluation metrics.