What is OCR / document scanning?
Optical Character Recognition (OCR) is the process of extracting machine-readable text from images or scanned documents — turning a photo of an invoice into structured data your app can use. When a user photographs a paper form, uploads a scanned PDF, or snaps an ID document, OCR reads and returns the text automatically.
Modern OCR goes beyond simple character recognition. AI-enhanced services like AWS Textract, Google Cloud Vision, and Azure Form Recognizer can identify document structure — tables, form fields, key-value pairs — not just raw text. This means "Invoice Total: $1,250.00" can be extracted as a labelled field rather than a line of unstructured text.
The accuracy of OCR depends heavily on document quality. A clean, high-contrast printed document scanned at 300 DPI will achieve near-perfect accuracy. A handwritten form photographed at an angle under fluorescent lighting is significantly harder — even the best AI models make errors. Your implementation needs to account for this through confidence scoring and human review queues.
When does your app need it?
- You want users to upload paper forms or scanned PDFs without manually re-typing data
- You're processing incoming invoices, receipts, or purchase orders at volume
- Your workflow involves ID document verification — driver's licences, passports, Medicare cards
- You need to digitise a backlog of historical paper records
- Staff currently re-key data from printed reports or faxed documents
- You're building an expense management feature where users photograph receipts
How much does it cost?
Adding OCR / document scanning typically adds 8–16 hours of development — roughly $1,000–$4,000 AUD.
The simpler end covers basic text extraction from clean, typed documents using a managed cloud API. The higher end involves building a full processing pipeline: upload handling, image pre-processing (deskew, contrast adjustment), OCR extraction, confidence scoring, structured data parsing, and a review queue for low-confidence results.
Ongoing API costs vary by provider and volume — AWS Textract charges per page processed. For high-volume workflows, this can become a meaningful operational cost worth modelling upfront.
How it's typically built
Most implementations call a managed OCR API rather than running models locally. AWS Textract, Google Cloud Vision, and Azure Form Recognizer all offer REST APIs that accept an image or PDF and return extracted text, confidence scores, and (for structured documents) labelled fields and tables.
For open-source requirements, Tesseract is the standard choice — free to run on your own infrastructure, but requires more tuning and produces lower accuracy on complex layouts.
A typical pipeline: the user uploads a file → your app stores it in S3 or equivalent → sends it to the OCR API → receives structured results → maps fields to your data model → flags low-confidence results for human review. Pre-processing steps (image rotation correction, noise reduction) are often worth adding for document types photographed in the field.
Questions to ask your developer
- What document types and quality levels have you tested? OCR accuracy varies significantly — get clarity on expected accuracy for your specific documents.
- How are low-confidence results handled? There should be a defined threshold and a review workflow, not silent errors.
- Which OCR provider is being used and why? AWS Textract is strong for structured forms; Google Vision is better for handwriting; Tesseract suits on-premise requirements.
- What's the per-page API cost at your expected volume? At high volume, OCR API costs can rival development cost within a year.
- Is pre-processing included? Images uploaded from mobile cameras often need rotation correction and contrast adjustment before OCR runs reliably.
See also: Document management · File versioning and history · App cost calculator