Implement computer vision solutions
Choose the right capability
Azure AI Vision image analysis returns captions, dense captions, tags, objects with bounding boxes, people, and Read-based OCR. Custom Vision (or Foundry custom models) trains on your own images: image classification gives one or more labels per image; object detection returns labels plus bounding boxes and can count instances. Document Intelligence handles forms and documents: prebuilt models (read, layout, invoice, receipt, ID) and custom models (template for stable layouts, neural for varied layouts, composed to route across several).
Matching model to requirement
Ask what the output must be. A single caption for a photo is image analysis. Sorting photos into damaged/undamaged is classification. Counting products on a shelf or finding their position is detection. Extracting line items and totals from a scanned invoice is Document Intelligence, not OCR.
Traps: using classification when images contain multiple objects; expecting plain OCR to return tables and key/value structure (that is the layout model); assuming the Face service is open to any caller - face identification is restricted-access and requires registration; forgetting custom models are region- and endpoint-bound, so the prediction resource must match.
Operational notes
Train with balanced, representative images and hold out a test set; watch precision vs recall depending on whether false positives or misses cost more. Use the Read model for printed and handwritten text, Video Indexer for media insights, and spatial analysis for edge camera analytics.
Worked rule: if the answer needs a count or a location, it is object detection; if it needs a single label for the whole image, it is classification.