How do I get key-value pairs out of invoices with OCR?

To extract key-value pairs from invoices with OCR, you usually need two steps:

1. OCR text extraction
   Convert the invoice image/PDF into machine-readable text.
2. Field extraction / key-value parsing
   Map the OCR text into structured fields like invoice_number, date, vendor_name, total, etc.

Common approaches

1) Rule-based parsing

Best if your invoices are from a few known vendors/templates.

• Use OCR to get text
• Search for labels like:
  • Invoice No:
  • Date:
  • Total:
• Extract values using regex or position-based rules

Pros: simple, cheap, fast
Cons: brittle across varying invoice layouts

---

2) Template-based extraction

Best when invoice formats are consistent.

• Detect layout regions
• OCR each region
• Associate labels with nearby values based on coordinates

This works well if you know the invoice structure in advance.

---

3) ML / document AI extraction

Best for many vendors and varying layouts.

Use specialized document understanding tools/models that do:

• OCR
• layout analysis
• key-value extraction

Examples:

• Amazon Textract
• Google Document AI
• Azure Form Recognizer / Document Intelligence
• Open-source models like LayoutLM, Donut, or custom pipelines

Pros: handles many formats better
Cons: more setup, training, or cost

---

Typical pipeline

1. Preprocess image

   • deskew
   • denoise
   • increase contrast

2. Run OCR

   • Tesseract, PaddleOCR, EasyOCR, cloud OCR, etc.

3. Extract text + coordinates

   • You want both the words and their bounding boxes

4. Find labels and values

   • Example: if Invoice Number: appears, capture text to the right of it
   • If Total appears near a currency amount, pair them

5. Normalize output

   • Convert to JSON:

   {
     "invoice_number": "INV-10293",
     "invoice_date": "2026-10-01",
     "vendor": "Acme Supplies",
     "total": "$1,245.50"
   }
   

---

If you’re using OCR text only

You can extract key-value pairs with regex patterns such as:

import re

text = """
Invoice No: INV-10293
Date: 2026-10-01
Total: $1,245.50
"""

invoice_no = re.search(r'Invoice\s*No[:\s]+(\S+)', text)
date = re.search(r'Date[:\s]+([0-9\-\/]+)', text)
total = re.search(r'Total[:\s]+\$?([0-9,]+\.\d{2})', text)

print({
    "invoice_number": invoice_no.group(1) if invoice_no else None,
    "date": date.group(1) if date else None,
    "total": total.group(1) if total else None
})

---

If you want better accuracy

Use OCR with bounding boxes. Then apply heuristics like:

• value is to the right of the label
• value is on the next line
• use the nearest numeric text for totals
• use layout-aware models for complex invoices

---

Recommended tools

• Simple OCR: Tesseract, EasyOCR, PaddleOCR
• Best out-of-the-box invoice extraction: AWS Textract, Google Document AI, Azure Document Intelligence
• Open-source advanced: LayoutLM, Donut

If you want, I can show you:

1. a Python example using Tesseract,
2. a cloud API approach, or
3. a layout-aware ML approach for invoices.