Skip to content

Latest commit

 

History

History
125 lines (75 loc) · 9.05 KB

README.md

File metadata and controls

125 lines (75 loc) · 9.05 KB
Raw

Image Pre-Processing

Before sending the source images to Textract, we may wish to:

  1. Validate their quality (or else reject early in the pipeline) or otherwise classify the type of document being received
  2. Enhance the image for OCR to mitigate minor quality issues
  3. Extract particular regions of interest, to remove other noise

In this example we demonstrate (1) with a small example dataset, and show some architectural tips for (2) and (3).

Component Architecture

With Amazon Rekognition Custom Labels, we can train and deploy an image classification model (for validation or classification) or bounding-box object detection model (could be useful for region-of-interest extraction) - with no ML experience required: Just labelled data.

For basic image processing techniques (for enhancement), it may be sufficient to install OpenCV directly in our AWS Lambda runtime and process the image there.

In our toy example the annotation is simple (classification, rather than bounding box), small (very few samples), and already done (images sorted into folders): But we'll show how you can take advantage of Amazon SageMaker Ground Truth to accelerate real-world data labelling tasks.

To maintain flexibility on whether the component simply validates images or in some way modifies them, we have our processing function return a (potentially changed) S3 URI - rather than just a label.

The resulting architecture is as shown below:


alt text

Post-Deployment Setup

To set up this module after deploying the solution, you'll need to:

  1. Fetch the training data (or bring your own!) into your environment
  2. (Optionally) Label the images as "good" or "bad"
  3. Train a Rekognition Custom Labels solution
  4. Deploy the model: connecting it to the demo solution stack

For a largely automated guide to this process via the AWS SDKs, see the Automated Setup.ipynb notebook in this folder. Otherwise, you can follow along with the instructions below:

1. Fetch the Data

Locate your training data bucket:

  • From the S3 Console, find the PreprocessTrainingBucket created for you by the CloudFormation stack: Its name should include this phrase, but will have some auto-generated prefix and suffix from CloudFormation/Amplify.
  • If you're not able to find the bucket in S3, go to your CloudFormation Stack and search in the "Resources" tab.
  • If you created the stack via AWS Amplify and don't know your CloudFormation Stack Name, it will be something like amplify-{AppName}-{BranchName}-{RandomNumber}-processing-{BranchName}

Download the sample data:

Download the sample (approx 100MB) receipts.zip from whichever of our S3 hosts is most convenient for your region:

Region URL
US East (N. Virginia) https://public-asean-textract-demo-us-east-1.s3.amazonaws.com/receipts.zip
Asia Pacific (Singapore) https://public-asean-textract-demo-ap-southeast-1.s3-ap-southeast-1.amazonaws.com/receipts.zip

Load the data into your bucket

You can simply extract receipts.zip on your local computer and then drag & drop the resulting good/ and bad/ folders into your PreprocessTrainingBucket via the S3 Console.

2. (Optionally) Label the Data

Since we have a toy, pre-labelled dataset for you, it's possible to skip SageMaker Ground Truth and import the S3 dataset directly to Rekognition Custom Labels.

To annotate the images in SageMaker Ground Truth:

  • Follow the instructions in groundtruth-labelling to annotate your images.
  • Open the Amazon Rekognition Custom Labels Console and navigate to Datasets from the left sidebar (which might be collapsed at first, if you're new to the service).
  • Click Create dataset and select the option to Import images labelled by Amazon SageMaker Ground Truth. You'll need to give the location of your output manifest file as created by the completed labelling job.

You shouldn't need to do any additional S3 bucket permissions configuration (if prompted), because this is set up by the CloudFormation stack. The dataset creation should look something like this:

Rekognition Create Dataset (Ground Truth) Screenshot

To skip data annotation and import directly:

  • Open the Amazon Rekognition Custom Labels Console and navigate to Datasets from the left sidebar (which might be collapsed at first, if you're new to the service).
  • Click Create dataset and select the option to Import images from an Amazon S3 bucket. You'll need to give the location of your output manifest file as created by the completed labelling job.
  • Provide the S3 location where your good and bad folders are uploaded.
  • Enable automatic labelling so your images will be automatically classified from their folder name.

Rekognition Create Dataset (S3) Screenshot

Whichever method you chose, you should now see a dataset in the Rekognition Custom Labels Console 'Datasets' list.

3. Train the Model

  • Go to the Projects List of the Amazon Rekognition Custom Labels Console and click Create project: Select a name that works for your context, e.g. demo-receipts-classifier.
  • Either from the main page or after clicking on the project, click "Train new model".
  • The correct Project ARN should have pre-populated for you - otherwise you can select a different project by name using the drop-down.
  • In Choose training dataset, select the Custom Labels dataset we just created above
  • In Create test set, you can either:
    • Apply good ML practice by selecting "Split training dataset" (since we don't have a separate test set defined in our example), or
    • Maximise performance on the (tiny) provided dataset by selecting "Choose an existing test dataset" and picking the same dataset again. In this case the model performance metrics won't be very meaningful, but our dataset is so small we know generalization won't be amazing anyway!

Rekognition Train Model Screenshot

Start the training process, and refresh the console periodically to check on the status of the job.

Start the training process, and then you'll need to wait until the training completes. When the model reaches TRAINING_COMPLETED status, it's ready to deploy. If you selected to use the same dataset for training and test, you'll likely see a reported F1 score of 1.000 - but if you chose to split the dataset you may see a significantly lower (and more representative!) figure.

Rekognition Model List Screenshot

3. Deploy and Integrate

Once your model has finished training, click on its name to view its full details page and copy the model ARN, which should have a format something like:

arn:aws:rekognition:us-east-1:123456789012:project/receipt-classification/version/receipt-classification.2020-08-03T15.57.13/1596441361045

Next, in the AWS SSM Parameter Store console, find the deployed stack's RekognitionModelArn parameter.

Edit your parameter to set the Value as your model ARN.

Your workflow should now be connected and ready to use the model for incoming requests!

You can start (deploy) and stop (de-provision) your Rekognition Custom Labels models through the Rekognition Custom Labels console. If the provided model is not already RUNNING, the Lambda function will request it to start when called.

  • This means the first invocation will fail, but kick off the deployment
  • Any subsequent invocations while the model is still STARTING will also return an error
  • ...But once the model is RUNNING it should work fine!

⚠️ Note: you're charged by provisioned capacity for the time that a Rekognition Custom Labels model is deployed, regardless of whether any requests are received.

To clean up resources when you're done experimenting, you can use the aws rekognition stop-project-version AWS CLI command or STOP your models through the Rekognition Custom Labels console.