Skip to content

🎨 Image Classification Solution (Azure AI Custom Vision)

🔍 Problem

You need to build a computer vision solution that can classify images into categories (e.g., identify products in a grocery checkout, classify fruit images) using your own training data.

☁️ Solution with Azure

Use Azure AI Custom Vision, which enables:

  • 📊 Image classification: Predict a single class (multiclass) or multiple classes (multilabel) for an image
  • 🎯 Object detection (not detailed in this material)

You train your own model using labeled images, then publish the model for prediction.

🧩 Components Required

  • 🏋️ Azure AI Custom Vision training resource Used to train custom models with your image data

  • 🔮 Azure AI Custom Vision prediction resource Used to generate predictions from new images using your trained model

  • 🌐 Azure AI Custom Vision portal Web interface at https://www.customvision.ai for managing projects, uploading images, training, publishing, and testing models

  • 📱 SDK / REST API

    • Python: azure-cognitiveservices-vision-customvision
    • .NET: Microsoft.Azure.CognitiveServices.Vision.CustomVision.Training and ...Prediction

  • 🆔 Project ID + resource endpoints/keys Required for training and prediction API calls

🏗️ Architecture / Development

🏋️ Training a model

  • Create an image classification project in the Custom Vision portal
  • Associate the project with a training resource
  • Upload images and assign class label tags
  • Train and evaluate the model in the portal
  • Publish the trained model to a prediction resource

💡 Both multiclass (single label per image) and multilabel (multiple labels per image) classification are supported.

🔮 Example: Classify images (Python SDK)

from msrest.authentication import ApiKeyCredentials
from azure.cognitiveservices.vision.customvision.prediction import CustomVisionPredictionClient

# Authenticate a client for the prediction API
credentials = ApiKeyCredentials(in_headers={"Prediction-key": "<YOUR_PREDICTION_RESOURCE_KEY>"})
prediction_client = CustomVisionPredictionClient(
    endpoint="<YOUR_PREDICTION_RESOURCE_ENDPOINT>",
    credentials=credentials
)

# Get classification predictions for an image
image_data = open("<PATH_TO_IMAGE_FILE>", "rb").read()
results = prediction_client.classify_image(
    "<YOUR_PROJECT_ID>",
    "<YOUR_PUBLISHED_MODEL_NAME>",
    image_data
)

# Process predictions
for prediction in results.predictions:
    if prediction.probability > 0.5:
        print(image, ': {} ({:.0%})'.format(prediction.tag_name, prediction.probability))

⭐ Best Practice / Considerations

🔄 Separate training and prediction resources

Enables flexibility (e.g., train in one region, deploy prediction resources in multiple regions)

🚀 DevOps integration

Use SDKs / REST API to automate training and publishing as part of a CI/CD pipeline

🛠️ Resource management

  • Each resource has its own endpoint and authentication key
  • Each project has a unique project ID used by client applications

📝 Sample Exam-Like Questions

1️⃣ What is required to train and publish an image classification model with Azure AI Custom Vision?

✅ Answer: A training resource for model training and a prediction resource for generating predictions

2️⃣ How can you classify images using a trained Custom Vision model?

✅ Answer: Submit images to the prediction resource using the SDK, REST API, or Custom Vision portal

3️⃣ What type of image classification is supported by Azure AI Custom Vision?

✅ Answer: Multiclass (one label per image) and multilabel (multiple labels per image) classification

4️⃣ What advantage does separating training and prediction resources provide?

✅ Answer: Flexibility to train in one region and deploy prediction endpoints in others for performance or compliance reasons

5️⃣ Which tool allows you to visually upload images, tag them, train, and test models?

✅ Answer: The Azure AI Custom Vision portal at https://www.customvision.ai