PIX-TAB: An Efficient Pixel-Precise Table Structure Recognition Method Using Speculative Decoding and Region-Based Image Segmentation

Researchers at Samsung R&D Institute Ukraine have proposed the PIX-TAB method, which accurately recognizes complex table structures while operating quickly in on-device environments. This model draws inspiration from the MTL-TabNet architecture and provides pixel-level precise structure recognition.

As a core technology, it introduces Position-Aware Pixel-Precise (PAPP) Tokens that explicitly provide row and column position information, improving recognition accuracy for long and complex tables. In addition, it applies Speculative Decoding technology to increase decoding speed while maintaining recognition accuracy, improving device responsiveness.

Key features are as follows:

• RBIS (Region-Based Image Segmentation): Uses the flood fill technique to reliably detect table cells with clear boundaries.
• Flexible extensibility: New languages can be supported simply by swapping the OCR model, without modifying the core structure model.
• Dataset enhancement: The model was trained on over 1 million synthetic data samples generated based on Wikipedia tables, with increased structural diversity.
• New evaluation metrics: To overcome the limitations of existing TSR (Table Structure Recognition) measurement methods, TEDS struct100 and TEDS 100 metrics were introduced.