Idea
Unified table recognition model delivering accurate structure, content, and layout extraction from images for streamlined data workflows.
Research Paper
Core Innovation
This paper introduces TableSeq, which formulates table structure recognition, content recognition, and cell localization as a single sequence generation task. Unlike prior multi-stage or multi-decoder approaches, it uses one decoder to produce interleaved HTML tags, text, and coordinate tokens, aligning logical, textual, and geometric information in one autoregressive sequence. This design simplifies architecture and post-processing while achieving competitive or superior accuracy.
Why It Matters
Accurate extraction of table data from images is critical for automating data entry, analysis, and integration across industries. TableSeq reduces reliance on multiple tools and complex pipelines, improving efficiency and scalability in document processing workflows. Its unified approach simplifies deployment and maintenance while maintaining high accuracy on diverse table layouts.
Market Size (TAM)
$2–10B TAM for document AI and table extraction; $500M–$1B SAM from enterprises and research institutions. Driven by growing digitization and automation of document workflows.
Potential Customers & Pain Points
- Financial institutions – Need precise extraction of tabular data from reports
- Research organizations – Require automated processing of scientific tables
- Enterprises with document workflows – Face complexity and errors from multi-stage table recognition pipelines
- Software developers – Seek compact end-to-end models for table data extraction.
Business Model
Offer TableSeq as a cloud API and on-premise SDK for document AI platforms, with tiered pricing based on usage and enterprise features. Provide customization and integration services for large clients.
Competitive Landscape
- TabFact
- PubTabNet models
- FinTabNet models
- SciTSR
- Google Document AI
Implementation Challenges
- Integration with existing enterprise document processing systems
- Handling extremely diverse and complex table layouts in real-world documents
- Scaling inference speed for large-scale deployment
Validation Strategy
- Benchmark against standard datasets (PubTabNet
- FinTabNet
- SciTSR) to confirm accuracy
- Pilot deployments with financial and research organizations to test real-world performance
- Measure inference speed and scalability in production environments
- Collect user feedback to refine model and API usability
Research Paper Overview
TableSeq: Unified Generation of Structure, Content, and Layout
Summary
TableSeq is an end-to-end image-only framework that jointly recognizes table structure, content, and cell location by generating a unified sequence of HTML tags, text, and coordinates. It eliminates the need for external OCR and complex post-processing, achieving state-of-the-art accuracy on multiple benchmarks with a compact architecture. The model also supports index-based table querying and faster decoding with minimal accuracy loss.