How do I extract key-value pairs and tables from invoices with OCR?

To extract key-value pairs and tables from invoices using OCR, the usual pipeline is:

1. Get the text from the invoice image/PDF
2. Detect layout structure (headers, rows, columns, fields)
3. Map text into fields like invoice number, date, total
4. Parse line-item tables into structured rows

1) OCR the document

Use an OCR engine to convert the invoice into text, ideally with bounding boxes for each word/line.

Common options:

• Tesseract: open source, good for basic OCR
• Google Document AI / Vision OCR
• Azure Form Recognizer
• AWS Textract
• PaddleOCR
• EasyOCR

For invoices, managed APIs often work better because they also return:

• text
• coordinates
• confidence
• tables/form structures

2) Extract key-value pairs

Invoices usually have labels like:

• Invoice Number: 12345
• Date: 2026-10-11
• Total: $1,234.56

Typical ways to extract them:

A. Rule-based matching

If the invoice formats are fairly consistent:

• search for keywords like Invoice No, Invoice #, Date, Total
• grab the nearby text to the right or below

Useful when:

• you have a few vendor templates
• layout is stable

B. Spatial pairing

Using OCR bounding boxes:

• find label words
• locate the closest value horizontally or vertically
• pair them based on position

Example:

• label at (x1, y1)
• nearest text block to the right on same line = value

C. ML/document AI extraction

Use invoice-specific extraction models that identify fields directly:

• invoice number
• vendor name
• subtotal
• tax
• total
• due date

This is better for:

• many vendors
• varied layouts
• noisy scans

3) Extract tables

Invoice line items usually appear in tables with columns like:

• Description
• Quantity
• Unit Price
• Amount

Approach options

A. Use a table extraction tool

Tools like:

• AWS Textract
• Azure Form Recognizer
• Google Document AI can detect tables and rows automatically.

B. Use OCR + layout analysis

If you only have OCR text and coordinates:

1. Detect column boundaries
2. Group words by row based on y-coordinate
3. Assign each word to a column by x-coordinate
4. Build rows into records

C. Use PDF table extractors

If invoices are PDFs with embedded text, try:

• Camelot
• Tabula
• pdfplumber These work best on digital PDFs, not scanned images.

4) Recommended practical pipeline

A strong real-world pipeline looks like:

• If PDF has text layer: parse text directly first
• If scanned/image: OCR it
• Run layout/table detection
• Extract key-value fields
• Extract line-item tables
• Normalize outputs into JSON

Example JSON:

{
  "invoice_number": "INV-10293",
  "invoice_date": "2026-10-11",
  "vendor_name": "ABC Supplies",
  "total": 1234.56,
  "line_items": [
    {
      "description": "Paper A4",
      "quantity": 10,
      "unit_price": 5.0,
      "amount": 50.0
    }
  ]
}

5) Tips for better accuracy

• Preprocess images: deskew, denoise, increase contrast
• Use high-resolution scans
• Crop headers/footers if they confuse OCR
• Standardize currency/date formats
• Validate totals: subtotal + tax = total
• Use confidence thresholds and manual review for low-confidence fields

6) If you want code

A simple Python approach could be:

• OCR with PaddleOCR or Tesseract
• Parse positions
• Apply heuristics for labels/values
• Use Camelot or Tabula for PDFs with tables

If you want, I can give you:

1. a Python example using Tesseract/PaddleOCR,
2. a cloud API approach with Textract/Azure/Google, or
3. a table + key-value extraction pipeline with sample code.