Custom Vision

Custom Vision is a powerful AI service, part of the Microsoft Azure Cognitive Services suite, designed to make computer vision accessible to everyone. It empowers developers and businesses to easily build, deploy, and refine custom models for image classification and object detection. Instead of requiring vast datasets and deep machine learning knowledge, Custom Vision utilizes transfer learning to create high-quality models from a small set of user-provided images. This significantly accelerates the development cycle, allowing you to go from idea to a production-ready model in a fraction of the time.

The platform is engineered for simplicity and efficiency. You can train a model to recognize specific visual concepts, such as identifying product types, detecting defects in manufacturing, or moderating visual content, all through an intuitive web-based interface. Once your model is trained, it can be seamlessly integrated into your applications via a simple REST API call or exported to run on edge devices for real-time, low-latency scenarios.

How to use Custom Vision

The process of creating a custom model is streamlined into a few simple steps:

1. Create a Project: Sign in to the Custom Vision portal, create a new project, and choose your project type—either Image Classification (to assign tags to an entire image) or Object Detection (to identify and locate specific objects within an image). You can also select a domain (e.g., Retail, Food, General) to optimize the model for your specific data.
2. Upload and Tag Images: Upload a set of training images. For classification, you apply one or more tags to each image. For object detection, you draw bounding boxes around the objects of interest in each image and assign a tag. The platform only requires a small number of examples per tag to get started (as few as 15).
3. Train the Model: Once your images are tagged, simply click the 'Train' button. Custom Vision's AI will analyze your images and build a custom model tailored to your specified concepts. The training process is fast and automated.
4. Evaluate and Iterate: After training, the service provides performance metrics like precision, recall, and mAP to help you understand your model's accuracy. You can use the 'Quick Test' feature to try out your model with new images. To improve performance, you can add more tagged images and retrain the model.
5. Deploy and Predict: When you're satisfied with the model's performance, you can publish it to get a prediction endpoint. Use the provided REST API to send new images to your model and receive predictions in a structured JSON format. Alternatively, you can export the model to run on various platforms, including iOS (CoreML), Android (TensorFlow), and Windows (ONNX), or as a Docker container for edge deployment.

Core Features of Custom Vision

• No-Code/Low-Code Interface: An intuitive graphical interface that guides you through the process of uploading data, training, and testing models without writing any machine learning code.
• Image Classification & Object Detection: Supports the two most common computer vision tasks, allowing you to classify entire images or locate multiple objects within them.
• Transfer Learning: Leverages pre-trained models from Microsoft, enabling you to build accurate custom models with very small datasets.
• Optimized Domains: Offers pre-optimized models for specific scenarios like retail, food, and landmarks to achieve higher accuracy out-of-the-box.
• REST API Endpoint: Easily integrate your custom model into any application with a simple and scalable cloud-based API.
• Edge Deployment: Export trained models to run on-device for applications requiring low latency, offline capabilities, or data privacy. Supported formats include CoreML, TensorFlow, ONNX, and Docker containers.
• Active Learning: The system can suggest which images you should label next to provide the most significant improvement to your model's performance, making the training process more efficient.

Use Cases for Custom Vision

Custom Vision is versatile and can be applied across numerous industries:

• Retail: Automated product recognition on shelves for inventory management, detecting empty shelf space, or powering visual search in e-commerce apps.
• Manufacturing: Quality assurance on production lines by automatically identifying defects, cracks, or other anomalies in parts.
• Healthcare: Assisting in the analysis of medical imagery to spot irregularities (for research and informational purposes).
• Agriculture: Monitoring crop health, identifying plant diseases, or counting produce for yield estimation.
• Security & Safety: Detecting safety equipment like hard hats on a construction site or identifying unauthorized objects in restricted areas.
• Digital Asset Management: Automatically tagging large volumes of images with custom labels for easier search and organization.

Advantages of Custom Vision

The primary advantage of Custom Vision is its ability to democratize computer vision. It lowers the barrier to entry, allowing teams without specialized AI talent to build powerful visual intelligence solutions. Its speed, efficiency with small datasets, and flexible deployment options (cloud and edge) make it a highly practical and cost-effective tool for a wide range of business problems.

Pricing and Plans

Custom Vision operates on a freemium model within the Azure ecosystem. A generous free tier is available, which typically includes a limited number of projects, training hours, and API prediction calls per month, making it ideal for learning, prototyping, and small-scale projects. For larger needs, the Standard (paid) tier follows a pay-as-you-go model. Costs are based on usage, including training time (per hour), the number of images stored, and the volume of prediction transactions. This flexible pricing ensures that you only pay for the resources you consume, making it scalable from small startups to large enterprises.