Share via

Issues with Signature Detection in Azure AI Document Intelligence Studio's Custom Neural Models

Forrest Middleton 0 Reputation points
2025-03-18T22:27:10.25+00:00

I'm encountering challenges with signature detection while training custom neural models in Azure AI Document Intelligence Studio. Despite following the recommended procedures, the models fail to recognize signature fields during analysis. Here's an overview of my approach:​

  • Sample Preparation: I've curated a dataset of over 20 documents, each labeled with signature fields using the Signature field type, ensuring that each field contains a single drawn region. There are up to 3 different signatures that could be detected, at least 5 samples are provided for each.​
  • Training Process: The model is trained using API version 2024-11-30 (GA), which, as per the official documentation, supports signature detection.​
  • Testing Outcomes: Post-training, when analyzing new documents, the model successfully identifies other labeled fields but consistently omits the signature fields. The fields are not visible in the RAW JSON and are not present in the response at all. Other fields that are not identified are shown with low confidence or are missing.

To troubleshoot, I've ensured that the training samples are diverse, the signature fields are labeled correctly, and the documents used for testing closely resemble the training set. Additionally, I've reviewed the raw JSON output to check for any low-confidence detections of signature fields but found none.​

I came across a similar issue, but the provided solutions haven't resolved my problem.​

Has anyone else experienced this issue or can provide insights into potential missteps or overlooked configurations? Any assistance or recommendations would be greatly appreciated.​

Thank you in advance for your support!

Azure AI Document Intelligence
0 comments

2 answers

Sort by: Most helpful
Most helpful Newest Oldest
  1. Anonymous
    2025-03-19T00:26:42.9733333+00:00

    Hi Forrest Middleton

    It sounds like you're experiencing an issue with signature detection in your custom neural models. Here are some potential areas to investigate based on the information available:

    • While you mentioned that you have over 20 documents with at least 5 samples for each signature, it's crucial to ensure that the samples are not only diverse but also representative of the variations in signatures you expect to encounter. If the training samples are too similar or do not cover the range of possible signatures, the model may struggle to generalize.
    • Double-check that the signature fields are labeled correctly and that each field contains only one drawn region. The model requires clear and precise labeling to learn effectively. If there are any inconsistencies in how the signature fields are labeled across your training documents, it could lead to issues during inference.
    • The custom neural model does have limitations, such as not recognizing values split across page boundaries. Ensure that the signatures are not located in a way that could cause them to be split across pages, as this could prevent detection.
    • Ensure that the documents you are testing closely resemble the training documents in terms of layout and signature placement. Significant differences in format or structure could affect the model's ability to detect signatures.
    • You are using the API version 2024-11-30 (GA), which supports signature detection. However, if there are any updates or patches related to signature detection, ensure that your implementation is up to date.
    • Since you mentioned that the signature fields are not visible in the RAW JSON output, it might be helpful to review the training process logs or any error messages that could provide insights into why these fields are not being detected.
    • kindly refer these document https://free.blessedness.top/en-us/azure/ai-services/document-intelligence/train/custom-neural?view=doc-intel-4.0.0#model-capabilities

    Hope this helps. Do let us know if you any further queries.

    If this answers your query, do click Accept Answer and Yes for was this answer helpful. And, if you have any further query do let us know.

  2. Van Bruwaene, W (Willem) [EXTERNAL] 0 Reputation points
    2025-10-24T09:03:22.3833333+00:00

    Hi, we encounter exactly the same issue. Is there already a solution in scope?

    BR

    0 comments

Your answer

Answers can be marked as 'Accepted' by the question author and 'Recommended' by moderators, which helps users know the answer solved the author's problem.