Document Intelligence custom classification model

Azure AI Document Intelligence is a cloud-based Azure AI service that enables you to build intelligent document processing solutions. Document Intelligence APIs analyze images, PDFs, and other document files to extract and detect various content, layout, style, and semantic elements.

Custom classification models are deep-learning-model types that combine layout and language features to accurately detect and identify documents you process within your application. Custom classification models perform classification of an input file one page at a time to identify the documents within and can also identify multiple documents or multiple instances of a single document within an input file.

Model capabilities

Custom classification models can analyze a single- or multi-file documents to identify if any of the trained document types are contained within an input file. Here are the currently supported scenarios:

• A single file containing one document type, such as a loan application form.

• A single file containing multiple document types. For instance, a loan application package that contains a loan application form, payslip, and bank statement.

• A single file containing multiple instances of the same document. For instance, a collection of scanned invoices.

✔️ Training a custom classifier requires at least two distinct classes and a minimum of five document samples per class. The model response contains the page ranges for each of the classes of documents identified.

✔️ The maximum allowed number of classes is 500. The maximum allowed number of document samples per class is 100.

The model classifies each page of the input document, unless specified, to one of the classes in the labeled dataset. You can specify the page numbers to analyze in the input document as well. To set the threshold for your application, use the confidence score from the response.

Incremental training

With custom models, you need to maintain access to the training dataset to update your classifier with new samples for an existing class, or add new classes. Classifier models now support incremental training where you can reference an existing classifier and append new samples for an existing class or add new classes with samples. Incremental training enables scenarios where data retention is a challenge and the classifier needs to be updated to align with changing business needs. Incremental training is supported with models trained with API version 2024-02-29-preview and later.

Incremental training requires that you provide the original model ID as the baseClassifierId. See incremental training to learn more about how to use incremental training.

Office document type support

You can now train classifiers to recognize document types in various formats including PDF, images, Word, PowerPoint, and Excel. When assembling your training dataset, you can add documents of any of the supported types. The classifier doesn't require you to explicitly label specific types. As a best practice, ensure your training dataset has at least one sample of each format to improve the overall accuracy of the model.

Compare custom classification and composed models

A custom classification model can replace a composed model in some scenarios but there are a few differences to be aware of:

Language support

Input requirements

Supported file formats:

• For best results, provide five clear photos or high-quality scans per document type.

• For PDF and TIFF, up to 2,000 pages can be processed (with a free tier subscription, only the first two pages are processed).

• The file size for analyzing documents is 500 MB for paid (S0) tier and 4 MB for free (F0) tier.

• Image dimensions must be between 50 x 50 pixels and 10,000 px x 10,000 pixels.

• If your PDFs are password-locked, you must remove the lock before submission.

• The minimum height of the text to be extracted is 12 pixels for a 1024 x 768 pixel image. This dimension corresponds to about 8-point text at 150 dots per inch (DPI).

• For custom model training, the maximum number of pages for training data is 500 for the custom template model and 50,000 for the custom neural model.

• For custom extraction model training, the total size of training data is 50 MB for template model and 1G-MB for the neural model.

• For custom classification model training, the total size of training data is 1 GB with a maximum of 10,000 pages.

Document splitting

When you have more than one document in a file, the classifier can identify the different document types contained within the input file. The classifier response contains the page ranges for each of the identified document types contained within a file. This response can include multiple instances of the same document type.

Best practices

Custom classification models require a minimum of five samples per class to train. If the classes are similar, adding extra training samples improves model accuracy.

The classifier attempts to assign each document to one of the classes, if you expect the model to see document types not in the classes that are part of the training dataset, you should plan to set a threshold on the classification score or add a few representative samples of the document types to an "other" class. Adding an "other" class ensures that unneeded documents don't affect your classifier quality.

Training a model

Custom classification models are supported by v4.0: 2024-02-29-preview, 2024-07-31-preview and v3.1: 2023-07-31 (GA) APIs. Document Intelligence Studio provides a no-code user interface to interactively train a custom classifier. Follow the how to guide to get started.

When using the REST API, if you organize your documents by folders, you can use the azureBlobSource property of the request to train a classification model.

https://{endpoint}/documentintelligence/documentClassifiers:build?api-version=2024-02-29-preview

{
  "classifierId": "demo2.1",
  "description": "",
  "docTypes": {
    "car-maint": {
        "azureBlobSource": {
            "containerUrl": "SAS URL to container",
            "prefix": "sample1/car-maint/"
            }
    },
    "cc-auth": {
        "azureBlobSource": {
            "containerUrl": "SAS URL to container",
            "prefix": "sample1/cc-auth/"
            }
    },
    "deed-of-trust": {
        "azureBlobSource": {
            "containerUrl": "SAS URL to container",
            "prefix": "sample1/deed-of-trust/"
            }
    }
  }
}

https://{endpoint}/formrecognizer/documentClassifiers:build?api-version=2023-07-31

{
  "classifierId": "demo2.1",
  "description": "",
  "docTypes": {
    "car-maint": {
        "azureBlobSource": {
            "containerUrl": "SAS URL to container",
            "prefix": "{path to dataset root}/car-maint/"
            }
    },
    "cc-auth": {
        "azureBlobSource": {
            "containerUrl": "SAS URL to container",
            "prefix": "{path to dataset root}/cc-auth/"
            }
    },
    "deed-of-trust": {
        "azureBlobSource": {
            "containerUrl": "SAS URL to container",
            "prefix": "{path to dataset root}/deed-of-trust/"
            }
    }
  }
}

Alternatively, if you have a flat list of files or only plan to use a few select files within each folder to train the model, you can use the azureBlobFileListSource property to train the model. This step requires a file list in JSON Lines format. For each class, add a new file with a list of files to be submitted for training.

{
  "classifierId": "demo2",
  "description": "",
  "docTypes": {
    "car-maint": {
      "azureBlobFileListSource": {
        "containerUrl": "SAS URL to container",
        "fileList": "{path to dataset root}/car-maint.jsonl"
      }
    },
    "cc-auth": {
      "azureBlobFileListSource": {
        "containerUrl": "SAS URL to container",
        "fileList": "{path to dataset root}/cc-auth.jsonl"
      }
    },
    "deed-of-trust": {
      "azureBlobFileListSource": {
        "containerUrl": "SAS URL to container",
        "fileList": "{path to dataset root}/deed-of-trust.jsonl"
      }
    }
  }
}

As an example, the file list car-maint.jsonl contains the following files.

{"file":"classifier/car-maint/Commercial Motor Vehicle - Adatum.pdf"}
{"file":"classifier/car-maint/Commercial Motor Vehicle - Fincher.pdf"}
{"file":"classifier/car-maint/Commercial Motor Vehicle - Lamna.pdf"}
{"file":"classifier/car-maint/Commercial Motor Vehicle - Liberty.pdf"}
{"file":"classifier/car-maint/Commercial Motor Vehicle - Trey.pdf"}

{
  "classifierId": "existingClassifierName",
  "allowOverwrite": true,  // Default=false
  ...
}

POST https://myendpoint.cognitiveservices.azure.com/documentintelligence/documentClassifiers:authorizeCopy?api-version=2024-07-31-preview
Ocp-Apim-Subscription-Key: {<your-key>}

{
  "classifierId": "targetClassifier",
  "description": "Target classifier description"
}

{
  "targetResourceId": "/subscriptions/targetSub/resourceGroups/targetRG/providers/Microsoft.CognitiveServices/accounts/targetService",
  "targetResourceRegion": "targetResourceRegion",
  "targetClassifierId": "targetClassifier",
  "targetClassifierLocation": "https://targetEndpoint.cognitiveservices.azure.com/documentintelligence/documentClassifiers/targetClassifier",
  "accessToken": "accessToken",
  "expirationDateTime": "timestamp"
}

POST {endpoint}/documentintelligence/documentClassifiers/{classifierId}:copyTo?api-version=2024-07-31-preview
Ocp-Apim-Subscription-Key: {<your-key>}

{
  "targetResourceId": "/subscriptions/targetSub/resourceGroups/targetRG/providers/Microsoft.CognitiveServices/accounts/targetService",
  "targetResourceRegion": "targetResourceRegion",
  "targetClassifierId": "targetClassifier",
  "targetClassifierLocation": "https://targetEndpoint.cognitiveservices.azure.com/documentintelligence/documentClassifiers/targetClassifier",
  "accessToken": "accessToken",
  "expirationDateTime": "timestamp"
}

Model response

Analyze an input file with the document classification model.

https://{endpoint}/documentintelligence/documentClassifiers/{classifier}:analyze?api-version=2024-02-29-preview

https://{service-endpoint}/formrecognizer/documentClassifiers/{classifier}:analyze?api-version=2023-07-31

The response contains the identified documents with the associated page ranges in the documents section of the response.

{
  ...

    "documents": [
      {
        "docType": "formA",
        "boundingRegions": [
          { "pageNumber": 1, "polygon": [...] },
          { "pageNumber": 2, "polygon": [...] }
        ],
        "confidence": 0.97,
        "spans": []
      },
      {
        "docType": "formB",
        "boundingRegions": [
          { "pageNumber": 3, "polygon": [...] }
        ],
        "confidence": 0.97,
        "spans": []
      }, ...
    ]
  }

Next steps

Learn to create custom classification models: