I'm building a mobile capture app and need OCR for bad photos

For bad photos, OCR works best when you don’t rely on OCR alone. The winning pattern is usually:

1. Detect/rectify the document or text region
2. Enhance the image
3. Run a robust OCR engine
4. Post-process and confidence-filter results

Best OCR options for low-quality mobile captures

Cloud OCR

Good if you want strong accuracy with minimal ML work.

• Google Cloud Vision OCR
  • Very solid on messy images
  • Good handwriting support relative to many others
  • Easy to integrate
• AWS Textract
  • Best for documents/forms/tables
  • Less ideal for general scene text
• Azure OCR / Document Intelligence
  • Strong document extraction
  • Good for business workflows

On-device OCR

Best if you need offline, privacy, or lower latency.

• Google ML Kit Text Recognition
  • Good mobile SDK
  • Easy on Android/iOS
  • Reasonable for casual OCR, less strong on very bad images
• Apple Vision framework
  • Excellent on iOS
  • Very efficient on-device
• Tesseract
  • Open source, flexible
  • Usually weaker on bad photos unless heavily preprocessed

Best accuracy for hard images

If you can use a custom pipeline/model:

• PaddleOCR
• EasyOCR
• Transformer-based OCR models for server-side inference

These can outperform traditional OCR on noisy text, especially if combined with preprocessing.

---

What to do for “bad photos”

If images are blurry, skewed, low light, or angled, add preprocessing:

• Auto-crop / document detection
• Perspective correction
• Deskew
• Deblur / sharpen
• Denoise
• Contrast / brightness normalization
• Adaptive thresholding
• Super-resolution for tiny text
• Text-region cropping before OCR

A simple pipeline often improves results a lot:

• OpenCV detect edges/corners
• Warp to flat document
• Convert to grayscale
• Denoise
• Adaptive threshold
• OCR

---

Recommended stack by use case

If you want easiest + best overall

• Google Cloud Vision OCR
• Add document detection and preprocessing on-device first

If you want offline mobile

• Apple Vision on iOS
• ML Kit on Android
• Add OpenCV preprocessing

If you want best control and can run backend inference

• PaddleOCR or a modern OCR model
• Use a preprocessing pipeline
• Add confidence-based rescans

---

Practical mobile app tips

• Ask users to capture with:
  • good lighting
  • focus lock
  • document filling most of the frame
  • avoid motion blur
• Use live edge detection and quality scoring before capture
• If confidence is low, automatically prompt for a retake
• Consider multi-shot capture and pick the sharpest frame

---

My recommendation

If your core problem is bad photos, I’d suggest:

• On-device preprocessing with OpenCV
• Cloud OCR for final extraction if privacy/latency allows
• Or Apple Vision / ML Kit if you need fully mobile

If you want, I can recommend a specific OCR stack for iOS, Android, or cross-platform, or give you a preprocessing pipeline in OpenCV.