NEW Fine-Tuning OpenAI Models: A Guide 🚀

Interactive bounding boxes OCR using Tesseract

Use an OCR engine for interactive ML-assisted labeling, facilitating faster annotation for layout detection, classification, and recognition models.

Tesseract is used for OCR but minimal adaptation is needed to connect other OCR engines or models.

Tested against Label Studio 1.10.1, with basic support for both Label Studio Local File Storage and S3-compatible storage, with a example data storage with MinIO.

Before you begin

Before you begin:

  • Ensure git is installed
  • Ensure Docker Compose is installed. For MacOS and Windows users, we suggest using Docker Desktop.

You must also install the Label Studio ML backend.

This tutorial uses the tesseract example.

1. Install Label Studio

Launch Label Studio. You can follow the guide from the official documentation or use the following commands:

docker run -it \
   -p 8080:8080 \
   -v `pwd`/mydata:/label-studio/data \
   heartexlabs/label-studio:latest

Optionally, you may enable local file serving in Label Studio

docker run -it \
   -p 8080:8080 \
   -v `pwd`/mydata:/label-studio/data \
   --env LABEL_STUDIO_LOCAL_FILES_SERVING_ENABLED=true \
   --env LABEL_STUDIO_LOCAL_FILES_DOCUMENT_ROOT=/label-studio/data/images \
   heartexlabs/label-studio:latest

If you’re using local file serving, be sure to get a copy of the API token from Label Studio to connect the model.

2. Create a Label Studio project

Create a new project for Tesseract OCR. In the project Settings set up the Labeling Interface.

Fill in the following template code. It’s important to specify smart="true" in RectangleLabels.

<View>    
   <Image name="image" value="$ocr" zoom="true" zoomControl="false"
         rotateControl="true" width="100%" height="100%"
         maxHeight="auto" maxWidth="auto"/>
   
   <RectangleLabels name="bbox" toName="image" strokeWidth="1" smart="true">
      <Label value="Label1" background="green"/>
      <Label value="Label2" background="blue"/>
      <Label value="Label3" background="red"/>
   </RectangleLabels>

   <TextArea name="transcription" toName="image" 
   editable="true" perRegion="true" required="false" 
   maxSubmissions="1" rows="5" placeholder="Recognized Text" 
   displayMode="region-list"/>
</View>

3. Install Tesseract OCR

Download the Label Studio Machine Learning backend repository.

git clone https://github.com/humansignal/label-studio-ml-backend
cd label-studio-ml-backend/label_studio_ml/examples/tesseract

Configure parameters in example.env file:

LABEL_STUDIO_HOST=http://host.docker.internal:8080
LABEL_STUDIO_ACCESS_TOKEN=<optional token for local file access>

AWS_ACCESS_KEY_ID=<set to MINIO_ROOT_USER for minio example>
AWS_SECRET_ACCESS_KEY=<set to MINIO_ROOT_PASSWORD for minio example>
AWS_ENDPOINT=http://host.docker.internal:9000

MINIO_ROOT_USER=<username>
MINIO_ROOT_PASSWORD=<password>
MINIO_API_CORS_ALLOW_ORIGIN=*

Depending on your data ingestion method, several configurations are possible:

Local File Storage

If you opted to use Label Studio Local File Storage, be sure to set the LABEL_STUDIO_HOST and LABEL_STUDIO_ACCESS_TOKEN variables.

S3-Compatible Storage (MinIO or AWS S3)

Configure the backend and the MinIO server by editing the MINIO_ROOT_USER AND MINIO_ROOT_PASSWORD variables, and make the AWS_ACCESS_KEY_ID AND AWS_SECRET_ACCESS_KEY variables equal to those values. You may optionally connect to your own AWS cloud storage by setting those variables. Note that you may need to make additional software changes to the tesseract.py file to match your particular infrastructure configuration.

Note: If you’re using this method, remove LABEL_STUDIO_ACCESS_TOKEN from the example.env file or leave it empty.

Other remote storage

If you host your images on any other public storage with http or https access, don’t change the default example.env file.

4. Start the Tesseract and MinIO servers

docker compose up

5. Upload tasks

If you’re using the Label Studio Local File Storage option, upload images directly to Label Studio using the Label Studio interface.

If you’re using MinIO for task storage, log into the MinIO control panel at http://localhost:9001. Create a new bucket, making a note of the name, and upload your tasks to MinIO.

Set the visibility of the tasks to be public. Further configuration of your cloud storage is beyond the scope of this tutorial, and you will want to configure your storage according to your particular needs.

If using MinIO, go to the Cloud storage page in the project settings.

Add your source S3 storage by connecting to the S3 Endpoint http://host.docker.internal:9000, using the bucket name from the previous step, and Access Key ID and Secret Access Key as configured in the previous steps. For the MinIO example, uncheck Use pre-signed URLS. Check the connection and save the storage.

6. Add model in project settings

From the project settings, select the Model page and click Connect Model.

Add the URL http://host.docker.internal:9090 and save the model as an ML backend.

7. Label in interactive mode

To use this functionality, activate Auto-Annotation and use the Autodetect rectangle for drawing boxes

Example below:

ls_demo_ocr

Reference links: