Handwritten Content and Signature Detection Using Azure AI Content Understanding

Hello Junaidh Ahamed,

Greetings! Thanks for raising this question in the Q&A forum.

Azure AI Content Understanding can be a reasonable tool for a "is handwriting present or not" check, but I cannot give you a fixed reliability figure for it. Results depend on your scan quality, the layout of your invoices and how you define the task, so it has to be measured on your own documents. Based on what the services document, here is how I would answer your three questions.

Handwritten content detection: when you only need to know whether handwriting exists, you do not need text recognition at all. The layout and read capabilities in Azure AI Document Intelligence, which Content Understanding builds on for documents, mark text lines that are in a handwritten style and give a confidence score for each, together with their position on the page. A simple rule such as "any handwritten line above a confidence threshold" gives you a yes or no. Content Understanding also lets you define your own analyzer with a field schema (for example a boolean field has_handwriting), and an AI model fills it in with a description that you write. That is flexible, but it is a generative approach, so results can vary more than a rule based on style confidence.

Signatures versus general handwriting: this is the hardest part. A signature is only a handwritten mark, and the handwritten style flag will not tell a signature from a handwritten note or a handwritten amount. Signatures are also often read as text or skipped by OCR when they are cursive or overlap printed lines. The documented signature detection feature in Document Intelligence is a signature field type in custom models, and it tells you whether a signature exists inside a specific labeled field. That works best when the signature sits in a predictable place, such as the "Authorized signature" box on an invoice. In Content Understanding you can describe the field in your schema ("a handwritten signature near the signature line, not general handwritten notes"), and the result can then be checked against your test set, but it is not a dedicated signature detector.

Known limitations to plan for: low resolution or faint scans, signatures that overlap printed text or stamps, handwriting in unsupported languages or unusual scripts, initials and checkmarks that look like signatures, very small signatures, and page layouts that differ from one vendor to another. You may also see service limits on file size, page count and supported regions, and Content Understanding features can change between preview and GA, so check the current documentation before you design around it.

Here are the steps I recommend:

Create a labeled sample. Gather 100 to 300 real invoices and business documents, and mark for each page whether it has handwriting, a signature, both or neither. Include poor scans, as they decide your real accuracy.

Start with the layout style signal for handwriting. Run the Document Intelligence layout or read model with the style add-on and check the styles collection for isHandwritten spans. Pick a confidence threshold, for example 0.7 or higher, and measure precision and recall on your sample.

https://<endpoint>/documentintelligence/documentModels/prebuilt-layout:analyze?api-version=<version>&features=styleFont

Test a Content Understanding analyzer in parallel. Define a schema with two boolean fields, has_handwriting and has_signature, add a clear description for each that separates a signature from general handwriting, and run the same sample through it. Compare it against step 2.

Add a signature field for fixed layouts. If most of your documents have a known signature area, train a custom Document Intelligence model with a signature field type in that area. Otherwise, for varied layouts, consider an object detection model trained in Azure AI Custom Vision or Azure Machine Learning on your own signature examples.

Combine the results. A common pattern is: handwritten style found in the signature region, or the custom signature field present, means signature. Handwritten style found elsewhere means handwritten content. Send low confidence cases to a person for review.

Improve scan quality. Use 200 to 300 DPI, keep colour, deskew and avoid heavy compression. This often improves results more than changing the model.

1. Re-measure regularly. Keep your labeled sample as a regression test, and run it again when the service version or your document sources change.If this answer helps you kindly accept the answer which will help others who have similar questions.

Best Regards,
Jerald Felix.