FastTab: Efficient Table Structure Recognition with Tiny Recursive Module and 1D Transformers

FastTab introduces a novel approach to table structure recognition (TSR), delivering impressive results across four benchmarks while maintaining low latency. This model sidesteps autoregressive HTML decoding through a grid-based methodology that integrates a Tiny Recursive Module (TRM) for comprehensive reasoning and axial 1D Transformer encoders to manage long-range dependencies in both rows and columns. It forecasts the number of rows and columns, identifies header rows, and determines separators to build a grid, subsequently deducing rowspan and colspan through ROI-aligned cell features. Evaluated on PubTabNet, FinTabNet, PubTables-1M, and SciTSR, FastTab demonstrates effective structure recovery. The research also investigates resilience against pixel-level anonymization and adapts to curved separators in documents captured by cameras. The source code will be available on GitHub.