This module is designed for engineers, data scientists, and architects who want to apply computer vision in production. Familiarity with basic machine learning concepts and common tooling (Python, ML frameworks, cloud APIs) will help you move faster, but is not strictly required.
- How pixel data is transformed into semantic, task-specific outputs (labels, bounding boxes, masks, text).
- The main model families used in vision: convolutional neural networks (CNNs), vision transformers (ViTs), diffusion models, and multimodal architectures.
- How Azure AI Vision and Microsoft Foundry map to typical image-analysis workflows and production needs.
- Practical trade-offs and when to choose prebuilt services, customizable models, or in-house training.
-
Describe core computer vision concepts and common image-analysis tasks
- Understand what computer vision does: representation of pixels, features, and semantics.
- Learn common tasks: image classification, object detection, semantic and instance segmentation, optical character recognition (OCR), pose estimation, object tracking, and image generation.
- Recognize evaluation metrics and typical datasets used to measure performance (accuracy, mAP, IoU, F1).
-
Differentiate between major types of computer vision models
- Compare convolutional neural networks (CNNs) and vision transformers (ViTs) for recognition and feature extraction.
- Understand diffusion models for image synthesis and generative tasks.
- Explore multimodal models that fuse image and text for retrieval, captioning, and visual question answering (VQA).
- Assess strengths, weaknesses, and pragmatic use cases for each family (e.g., latency vs. accuracy, data requirements, transfer learning potential).
-
Identify Azure AI Vision capabilities for image analysis workloads
- Map common computer vision tasks to Azure AI Vision features: prebuilt models, custom training, OCR, spatial analysis, and batch vs. real-time inference.
- Understand when to use managed APIs versus custom models or hybrid approaches (e.g., custom object detection on top of a prebuilt OCR pipeline).
-
Recognize how Microsoft Foundry supports computer vision, image generation, and multimodal AI solutions
- See how Foundry can integrate model outputs, manage data and labels, and orchestrate multimodal pipelines.
- Learn how these platform capabilities support training, inference, monitoring, and operationalization for production vision systems.
Implementation considerations
- Data: Quality labels and diverse samples significantly impact model generalization. Use annotation tooling and labeling audits.
- Compute: Training large vision models requires GPU/TPU resources; inference latency varies by model family and deployment target.
- Cost & compliance: Managed services reduce operational burden but can have different cost profiles and data residency considerations.
Some computer vision workflows (training, high-throughput inference, large-scale labeling) can incur significant cloud costs. Plan experiments and production deployments with budget and compliance constraints in mind.
- Conceptual foundations: core tasks, evaluation, and model families.
- Practical patterns: pipelines for labeling, training, inference, monitoring, and scaling.
- Platform alignment: mapping tasks to Azure AI Vision, Foundry, and relevant Azure services.
- Examples and walkthroughs: end-to-end examples that combine prebuilt services with custom models and orchestration.
- Kubernetes Documentation — for deploying scalable inference services
- Azure AI Vision Documentation — official Azure docs on vision services
- [Diffusion Models and Image Generation — Survey Papers] — for deeper theory on generative models