How to migrate stamp and signature detection from customvision.ai?

Question

How to migrate stamp and signature detection from customvision.ai?

Andrei Pruteanu 40

Hi,

We received a retirement notice related to customvision.ai:

Azure AI Custom Vision will be retired on 25 September 2028, please transition to alternative options by that date.

From now until 25 September 2028, you can continue to use Azure AI Custom Vision without disruption. However, customers are encouraged to begin planning and executing their transition to alternative solutions earlier to experience new capabilities.

Depending on your use case, we recommend the following paths for transition:

For creating custom models for both image classification and object detection, Azure Machine Learning AutoML offers the ability to train both custom model types using classic machine learning techniques.
Microsoft is also investing in Generative AI-based solutions that increase accuracy in custom scenarios using prompt engineering and other techniques.

To use generative models, you can use one of models available in the Azure AI Foundry model catalog and create your own solution for customized vision.
For a managed generative solution for image classification, Azure AI Content Understanding (currently in public preview) offers the ability to create custom classification workflows. It also supports processing unstructured data of any type (image, documents, audio, video) and extract structured insights based on pre-defined or user-defined formats.

What path do you recommend for migrating or creating new models for stamp and signature detection? We are currently using DocumentIntelligence Studio and customvision.ai.

Alex Burlachenko 18,310 Reputation points Volunteer Moderator

2025-10-03T08:14:42.4566667+00:00
hi Andrei Pruteanu,

you're using custom vision for a very specific task, stamp and signature detection, and now you need to find the best new home for it.

for your specific use case, the most straightforward and powerful path is to use azure ai document intelligence. you're already using the studio, which is perfect.

document intelligence has a dedicated 'custom neural' model type that is perfect for what you're doing. it's designed specifically for extracting custom information from documents, like finding stamps and signatures. you can train a custom model to recognize these elements with high accuracy, and it's built on a more modern architecture than the retiring custom vision service.

the migration process is relatively smooth. you can export your existing training data, your labeled images of stamps and signatures, from custom vision. then, you can import this data into a new custom model in document intelligence and retrain it. the underlying concept of labeling regions of interest is very similar.

the other options, like azure machine learning automl, would require more setup and infrastructure management. since you're already in the document intelligence ecosystem, leveraging its custom models is the most efficient path.

don't overcomplicate it. use the custom models within azure ai document intelligence to retrain your stamp and signature detector. it's the natural successor for your workflow.

regards,

Alex

and "yes" if you would follow me at Q&A - personaly thx. P.S. If my answer help to you, please Accept my answer

https://ctrlaltdel.blog/
Andrei Pruteanu 40 Reputation points

2025-10-03T20:24:56.2833333+00:00

Hi @Alex Burlachenko ,

Thank you for the answer. It sounds encouraging but do not know how to label stamps using Document Intelligence Studio / Custom Extraction Model. I only see the following: Field, Selection mark, Signature and Table (see attached image). Even more so, I can only label one signature per page using "Draw region". I have upto 4-5 signatures and stamps I need to label per page. customvision.ai had no issues whatsoever with this.

regards,
Andrei
SRILAKSHMI C 8,295 Reputation points Microsoft External Staff Moderator

2025-10-06T17:18:30.4866667+00:00

Hi Andrei Pruteanu,

Thank you for your follow-up, and I completely understand the confusion. This is a common scenario when moving from Custom Vision to Document Intelligence, as the labeling approach and feature set work a bit differently in the new platform.

The reason you only see options like Field, Selection Mark, Signature, and Table is that the built-in “Signature” type in Document Intelligence is treated as a semantic field, not as an object detector. Because of this, it currently supports labeling only one signature per page.

To work around this, you can label all stamps and signatures as “Field” regions instead of using the built-in “Signature” option. There is no restriction on how many Field labels you can have per page, so you can create multiple regions for example, Signature_1, Signature_2, Stamp_1, and so on. Later, in your processing logic, you can group or handle these fields as needed.

If your primary goal is to detect multiple visual objects like stamps, seals, or signatures across a document, the most effective approach is to use the Custom Neural model instead of the “Custom Extraction” type. The Custom Neural model is specifically designed for object-detection-style use cases and supports multiple bounding boxes per page.

This means you can label several instances of stamps or signatures on a single page, and the model will learn to detect them all independently similar to how Custom Vision worked.

A successful migration from Custom Vision to Document Intelligence follows these steps:

Export your labeled training images from Custom Vision.

Create a Custom Neural model project in Document Intelligence Studio.

Label each signature or stamp as a Field (or custom name) there is no limit per page.

Train and evaluate the model, which will now detect all instances automatically.

If you need not only to detect where the stamps or signatures are but also to extract specific information from them (for example, signer names or stamp types), you can adopt a two-step approach.

Train a Custom Neural model to detect all the elements. Then, use a Custom Extraction model or custom logic to extract the relevant metadata from the detected regions. This combination gives you the best of both worlds accurate detection and rich data extraction.

I Hope this helps. Do let me know if you have any further queries.

Thank you!

Answer accepted by question author

2 additional answers

Your answer

Alex Burlachenko 18,310 Reputation points Volunteer Moderator

2025-10-03T08:14:42.4566667+00:00

hi Andrei Pruteanu,

you're using custom vision for a very specific task, stamp and signature detection, and now you need to find the best new home for it.

for your specific use case, the most straightforward and powerful path is to use azure ai document intelligence. you're already using the studio, which is perfect.

document intelligence has a dedicated 'custom neural' model type that is perfect for what you're doing. it's designed specifically for extracting custom information from documents, like finding stamps and signatures. you can train a custom model to recognize these elements with high accuracy, and it's built on a more modern architecture than the retiring custom vision service.

the migration process is relatively smooth. you can export your existing training data, your labeled images of stamps and signatures, from custom vision. then, you can import this data into a new custom model in document intelligence and retrain it. the underlying concept of labeling regions of interest is very similar.

the other options, like azure machine learning automl, would require more setup and infrastructure management. since you're already in the document intelligence ecosystem, leveraging its custom models is the most efficient path.

don't overcomplicate it. use the custom models within azure ai document intelligence to retrain your stamp and signature detector. it's the natural successor for your workflow.

regards,

Alex

and "yes" if you would follow me at Q&A - personaly thx. P.S. If my answer help to you, please Accept my answer

https://ctrlaltdel.blog/
Andrei Pruteanu 40 Reputation points

2025-10-03T20:24:56.2833333+00:00

Hi @Alex Burlachenko ,

Thank you for the answer. It sounds encouraging but do not know how to label stamps using Document Intelligence Studio / Custom Extraction Model. I only see the following: Field, Selection mark, Signature and Table (see attached image). Even more so, I can only label one signature per page using "Draw region". I have upto 4-5 signatures and stamps I need to label per page. customvision.ai had no issues whatsoever with this.

regards,
Andrei

Answer 1

Hi Andrei Pruteanu,

I sincerely apologize for the delayed response.

Thank you for your question and for catching that detail, you are absolutely correct. The Custom Neural model is not currently visible or available in the Document Intelligence Studio interface. This capability is still in private/limited preview, meaning only select customers or internal testing environments have access to it at this time.

At present, the Document Intelligence Studio supports two main custom model types: Custom Extraction (for structured data field extraction) and Custom Classification (for document sorting). The Custom Extraction model can recognize and extract labeled fields, including signature areas, but it supports only one signature per page when using the built-in "Signature" label type. The object detection–style flexibility you had with Custom Vision (detecting multiple stamps or signatures per page) is not yet part of the Studio UI.

If your current goal is to detect multiple visual elements like stamps or signatures on a single page, there are two paths you can take. The first option is to continue using Custom Extraction models, but instead of the “Signature” label type, define multiple labeled fields such as Signature_1, Signature_2, Stamp_1, etc. This allows you to tag and train the model to recognize several regions per page. While this isn’t full object detection, it can still help identify and extract multiple regions effectively.

Another option is to use Azure AI Foundry or Azure Machine Learning if you need a more advanced object detection approach. You can train a model externally (for example, using a YOLO-based architecture) to detect stamps or signatures, and then integrate that detection output with Document Intelligence for text extraction. Azure AI Foundry makes this integration easier and provides flexibility similar to what you had with Custom Vision.

If you only need to identify multiple stamps or signatures on a page, the Custom Extraction workaround should meet your needs. However, if you require true object detection or more complex visual analysis, combining Azure AI Foundry or Azure ML with Document Intelligence is the better choice. This approach allows you to handle visual detection and text extraction together in one automated workflow.

Microsoft is actively working on making the Custom Neural model generally available. This new capability will bring back advanced object detection features similar to Custom Vision but optimized for document scenarios, enabling you to train models that detect multiple visual elements per page directly within Document Intelligence Studio.

Thank you!

Answer 2

Andrei Pruteanu 40

Hello SRILAKSHMI C,

How can I create a Custom Neural Model? I only see: Custom extraction model and Custom classification model (see attached picture).

Best,
Andrei
User's image

SRILAKSHMI C 8,295 Reputation points Microsoft External Staff Moderator

2025-10-08T18:35:23.45+00:00

Hi Andrei Pruteanu,

Thank you for reaching out. The reason you are only seeing the “Custom extraction model” and “Custom classification model” options in Document Intelligence Studio is because the “Custom Neural Model” feature is not yet generally available in the Studio interface.

Currently, the available options

Custom extraction model – used to label and train models to extract specific fields or data points from your documents.

Custom classification model – used to split and categorize documents into different classes based on their content.

The Custom Neural Model is a more advanced capability designed for scenarios where you want to build a neural network-based extraction model using your own labeled data. However, this feature is currently in limited or private preview and is not exposed in the Document Intelligence Studio UI for most users.

To work with custom neural models today, you typically need to use the SDK or REST API where the relevant training endpoints are available. Alternatively, organizations that are part of the private preview program may see this feature enabled in their workspace.

Thank you!
Andrei Pruteanu 40 Reputation points

2025-10-09T07:32:32.5833333+00:00

Hi SRILAKSHMI C,

That means your comment/suggestion:
"[..] If your primary goal is to detect multiple visual objects like stamps, seals, or signatures across a document, the most effective approach is to use the Custom Neural model instead of the “Custom Extraction” type. The Custom Neural model is specifically designed for object-detection-style use cases and supports multiple bounding boxes per page.[..]" does NOT apply in current version of Document Intelligence Studio. Is maybe Ai Foundry an option as well?

What do I now than? What is the suggested migration option from customvision.ai to Document Intelligence Studio? I find it absolutely strange to annotate documents via REST api (if that is available either).

best,
Andrei
SRILAKSHMI C 8,295 Reputation points Microsoft External Staff Moderator

2025-10-16T16:55:36.53+00:00

Hi Andrei Pruteanu,

Did you get any chance to review the above response. Do let me know if you have any further queries.

Thank you!

Answer 3

Andrei Pruteanu 40

Hi SRILAKSHMI C,

Thank you for the clarifications! I do not have any further queries at the moment.

Best,
Andrei

SRILAKSHMI C 8,295 Reputation points Microsoft External Staff Moderator

2025-10-17T11:24:35.73+00:00

Hi Andrei Pruteanu,

Thank you for the clarification, Since I’ve converted my earlier comment into an answer, could you please take a moment to mark it as Accepted? This helps others in the community with the same question find the solution more easily.

Share via

How to migrate stamp and signature detection from customvision.ai?

2 additional answers

Your answer