What it is
Amazon Textract is an AWS service that extracts text, handwriting and data from scanned documents and photos. Beyond plain OCR, it understands structure: forms as key and value pairs, tables as rows and columns, and specialized documents such as identity documents, invoices and receipts. It is called through an API and needs no model training.
Its AnalyzeExpense API is built for invoices and receipts. It returns the vendor, date, totals and each line item, and every field comes with a confidence score and its position on the page. Pricing is per page, so a single receipt costs about a cent. The confidence scores let an application decide which fields a person should check.
How I use it
- The receipt reader on mervinc.com sends a photo to AnalyzeExpense and gets back the vendor, date, every line and the total.
- Boxes are drawn on your own photo where each field was found. Point at a row to light up its box.
- The photo is shrunk to 1600 pixels in the browser before it leaves, sent to Textract and not stored anywhere.
- Five reads per visitor a day, counted in DynamoDB before Textract is called.
Textract is the right tool when people are copying numbers off paper into a system. The confidence score is the part worth designing around: a field at 100 needs a glance, and a lower one is the one to put in front of a person. That keeps people checking numbers instead of keying them in.
Questions
Can AWS read receipts automatically?
Yes. Textract's AnalyzeExpense returns the vendor, date, line items and total from a photo, each with a confidence score. You can try it on my receipt page.
Is the photo kept?
No. On my site the photo is shrunk in your browser, sent to Textract to be read and not stored anywhere.