Most businesses tend to have a long pipeline of administrative processes that are time and labor intensive. Azure’s Form Recognizer can automatically extract text, key-value pairs and tables from scanned documents without the need for extensive manual configuration of the service, to quickly get accurate results, with a fraction of the cost.
The service overcomes the configuration challenges of regular OCR (Optical Character Recognition) software by using machine learning to automatically “read” documents. The simplest way to start is by using its Layout API, that extracts text and table structures, while also giving the position of each piece of information by returning its bounding box coordinates.
It also offers the ability to train custom models, so that you can tailor the service specifically to your forms by using supervised or unsupervised learning, and it has a prebuilt receipt model using receipts from sales made in the USA.
With the prebuilt receipt model, by sending a picture of a wrinkled receipt:
The service returns the data contained in the document, in JSON format, automatically, and can be parsed to a table or any other type of representation.
| Merchant | Contoso |
| Address | 123 Main Street Redmond, WA 98052 |
| Phone number | +9876543210 |
| Date | 2019-06-10 |
| Time | 13:59:00 |
| Subtotal | 1098.99 |
| Tax | 104.4 |
| Total | 1203.39 |
By customizing the model, relationships between the information contained in the form can be created and extracted automatically by the service.
Currently on preview, it supports only the English language and has a great and extensive documentation on the official Microsoft website.
Form Recognizer can be accessed via REST API or by using the .NET SDK to perform information extraction and custom model training. For organizations that handle administrative processes manually, seek efficiency improvement and costs reduction, Form Recognizer can help to achieve those goals.
