handwriting OCR accuracy bad on receipts

Handwritten OCR on receipts is usually bad because receipts are a tough combination of:

• small, low-quality text
• crumpled/curved paper
• bad lighting/shadows
• mixed print + handwriting
• tight spacing and noise
• non-standard handwriting

To improve accuracy, try this:

1) Improve the image first

• Use higher resolution photos/scans
• Take the picture straight-on
• Use bright, even lighting
• Avoid blur; keep the phone steady
• Crop tightly to the receipt
• Flatten the receipt if possible
• Increase contrast and sharpness before OCR

2) Preprocess the image

Common steps that help:

• deskew the receipt
• denoise
• convert to grayscale
• binarize/adaptive threshold
• dewarp curved receipts

3) Use receipt-specific OCR

General OCR often struggles. Better options:

• Google Cloud Vision
• AWS Textract
• Azure Document Intelligence
• Receipt-focused ML models or fine-tuned OCR systems

4) Segment handwriting separately

If the receipt has printed text plus handwritten notes:

• detect printed text and handwriting as different regions
• OCR them with different models if possible

5) Use post-processing

• correct common character confusions: 0/O, 1/l/I, 5/S
• use context rules for totals, dates, and item names
• validate with business logic, e.g. totals matching line items

6) Fine-tune for your data

If this is a production problem:

• collect examples of your receipts
• fine-tune a handwriting OCR model on your domain
• include your most common handwriting styles

7) If handwriting is the main issue

Pure OCR may not be enough. Consider:

• a human-in-the-loop review step
• an LLM or custom parser to clean OCR output after recognition

If you want, I can help you with:

1. an image preprocessing pipeline for receipts,
2. which OCR engine to use, or
3. a Python/OpenCV example to improve accuracy.