I'm building a document parser and want OCR with tables and coordinates

For OCR with tables and coordinates, you’ll want a system that returns:

• Recognized text
• Bounding boxes / coordinates for each word, line, or block
• Table structure detection
• Ideally page-level layout information too

Good options

1) Cloud OCR APIs

Best if you want strong accuracy and less engineering.

• AWS Textract

  • OCR + tables + forms
  • Returns coordinates for words/lines/tables/cells
  • Good for invoices, forms, scanned docs

• Google Document AI

  • Strong layout and table extraction
  • Returns bounding boxes, tables, paragraphs, tokens
  • Good for general document parsing

• Azure Document Intelligence (Form Recognizer)

  • OCR + tables + key-value pairs
  • Gives polygons/coordinates for elements
  • Good enterprise option

2) Open-source stack

Best if you need self-hosting or customization.

• PaddleOCR

  • Strong OCR
  • Can return text boxes and confidence
  • Often paired with layout/table models

• Tesseract

  • Basic OCR with box coordinates
  • Weaker on complex layouts and tables

• LayoutParser + Detectron2 models

  • Useful for layout detection
  • Can identify tables, headers, paragraphs, etc.

• Table Transformer / PubTables-1M models

  • Good for table structure extraction
  • Often used with OCR output

Recommended approach

If your priority is tables + coordinates + production readiness, I’d suggest:

• AWS Textract if your docs are mostly forms/invoices and you want simplicity
• Google Document AI if you need rich layout/table structure
• Azure Document Intelligence if you’re already in Microsoft ecosystem

If you want open-source:

• Use PaddleOCR for OCR
• Use a table detection/structure model for table layout
• Merge OCR boxes with detected table cells

Data you should expect

A good OCR parser should output something like:

{
  "page": 1,
  "text": "Item Qty Price",
  "blocks": [
    {
      "type": "table",
      "bbox": [100, 200, 500, 600],
      "cells": [
        {"row": 1, "col": 1, "text": "Item", "bbox": [110, 210, 180, 240]},
        {"row": 1, "col": 2, "text": "Qty", "bbox": [190, 210, 250, 240]}
      ]
    }
  ]
}

If you want to build it yourself

A practical pipeline is:

1. Preprocess image/PDF

   • deskew
   • denoise
   • enhance contrast

2. Detect layout

   • text blocks
   • tables
   • figures

3. Run OCR

   • get words + boxes

4. Table reconstruction

   • assign OCR tokens to table cells
   • infer rows/columns from grid lines or spatial alignment

5. Export structured JSON

   • page, block, line, word, bbox, table cells

Quick recommendation

• Need fastest path: Textract / Document AI
• Need self-hosted: PaddleOCR + table detection model
• Need just OCR boxes: Tesseract or PaddleOCR
• Need strong table extraction: Document AI or Textract

If you want, I can also give you:

1. a Python example
2. a comparison table of OCR tools
3. a full JSON schema for OCR + tables + coordinates