Overview
This skill area covers three kinds of visual work. First, analyzing images with Azure Vision in Foundry Tools: choosing visual features such as captions, tags, objects, people, smart crops and Read (OCR, optical character recognition), sending them in one request, and interpreting the JSON that comes back, including printed and handwritten text. Second, custom image models: deciding between image classification and object detection, labeling images, training, reading precision, recall and mean average precision, then publishing and calling the model, including building it from code.
Third, video: Azure AI Video Indexer for transcripts, faces, labels, scenes and other insights from recorded or live video, and Spatial Analysis for counting and tracking the movement of people in camera feeds. The domain links to Domain 1 for service selection and containers and to Domain 6, where OCR and image analysis become skills in an Azure AI Search enrichment pipeline.
How it's tested
Questions describe an image or video requirement and ask which feature, project type or service to use, or show a request and JSON response to interpret. Expect code completions for the analyze call and the visual features list, metric interpretation for custom models, and steps in order for training and publishing a Custom Vision model.
-
Feature or service selection
A requirement such as alt text, product location on a shelf or handwritten notes is described and you choose the visual feature, the custom project type or the service.
-
Response and metric reading
You read a JSON response with bounding boxes and confidence scores, or a metrics panel with precision, recall and mAP, and decide what it means or what to change.
-
Train, publish and consume
You order the steps or complete code for creating tags, uploading labeled images, training, publishing an iteration and calling the prediction endpoint.
A retailer wants to know where each cereal box sits on shelf photos and how many of each brand appear. The prebuilt objects feature returns generic labels such as box, not brands. Image classification would say which brands are in the photo but not where or how many. A Custom Vision object detection project, trained with bounding boxes per brand, returns each box's location and tag. Location plus custom labels means custom object detection.
Skill groups 3
Official skill groups from the Microsoft outline, in outline order.
4.1 Analyze imagesVisual features, objects and tags, response interpretation, and printed and handwritten text extraction.
- What it means
- Image Analysis in Azure Vision takes an image URL or binary and a list of visual features (caption, dense captions, tags, objects, people, smart crops, read) and returns a JSON document with results per feature, including confidence scores and bounding boxes. The Read feature extracts printed and handwritten text as blocks, lines and words with bounding polygons.
- What to understand
- You request only the features you need, because each feature has its own processing and output. Captions describe the whole image in a sentence; dense captions describe up to ten regions; tags are keywords; objects add bounding boxes. Read handles mixed printed and handwritten text in one call; for multi-page PDFs and forms, Document Intelligence's read or layout model is the better fit. Caption features are offered only in some Azure regions, a detail that can decide a deployment question.
- How it shows up
- A code snippet sends an analyze request with a features list and you add the missing feature, or a JSON response is shown and you identify where the extracted text or object coordinates are.
- Common confusion
- Tags and objects get confused. Tags list what is in the image with no location; objects return a bounding box for each detected item.
4.2 Implement custom vision modelsClassification vs. object detection, labeling, training, metrics, publishing, consuming and code-first builds.
- What it means
- Azure Custom Vision trains image classification models (multiclass, one label per image, or multilabel, several labels per image) and object detection models (labels with bounding boxes). You label images, train an iteration, review precision, recall and average precision, publish the iteration to a prediction resource, and call it with the project ID and published name.
- What to understand
- Training and prediction use separate resources and keys. Precision is the share of predictions that were correct; recall is the share of real items that were found; raising the probability threshold usually increases precision and lowers recall. Object detection also reports mAP (mean average precision) and uses an overlap threshold for matching boxes. Compact domains allow export to ONNX, TensorFlow, Core ML or a Docker container for edge use. The code-first route uses the training SDK to create the project and tags, upload images with regions, train and publish.
- How it shows up
- Expect a choice of project type and domain from a requirement, a step order for train and publish, or a metric panel where you decide whether to change the threshold or add more images.
- Common confusion
- Calling the model with the training key and endpoint. Predictions require the prediction resource's key and endpoint and a published iteration name.
4.3 Analyze videosVideo Indexer insights from video or live streams, and Spatial Analysis for people presence and movement.
- What it means
- Azure AI Video Indexer runs many audio and visual models over a video and returns insights such as transcripts and translations, OCR text, labels, keywords, topics, named entities, scenes, shots and keyframes, and faces (a limited-access feature). Spatial Analysis processes live camera feeds at the edge to detect people's presence and movement, such as counting people in a zone or detecting line crossings and distances.
- What to understand
- Video Indexer is the choice when the requirement is searchable insights from video content, with a portal, API and embeddable player and insights widgets. Spatial Analysis is about where people are and how they move, not who they are; it runs as a container on an edge device with a GPU and emits events for operations such as person count, line crossing, zone crossing and distance.
- How it shows up
- A scenario asks for a searchable transcript and on-screen text from training videos (Video Indexer) or for occupancy alerts from store cameras (Spatial Analysis), and you choose the service or operation.
- Common confusion
- Assuming Spatial Analysis identifies individuals. It detects and tracks people anonymously for counting and movement, and face identification is a separate, limited-access capability.
High-yield concepts
The terms, rules and formulas this domain turns into questions most often.
-
Image Analysis visual features
The analyze call takes a list of features such as caption, denseCaptions, tags, objects, people, smartCrops and read, and returns a result section for each one requested.
- Remember
- Results only appear for requested features, so a missing section in the response usually means the feature was not in the request.
- Tested as
- Code completions ask which feature value produces the required output.
- Trap
- Requesting tags when the requirement needs object locations.
-
Caption vs. dense captions
Caption returns one sentence describing the whole image; dense captions return sentences for up to ten regions with bounding boxes.
- Remember
- Use gender-neutral captions when required; caption features are available only in certain regions.
- Tested as
- Accessibility requirements such as alt text point to caption.
- Trap
- Choosing dense captions for a single alt-text string.
-
Read OCR output
The Read feature returns text as blocks, lines and words, each with a bounding polygon and words with confidence scores.
- Remember
- It reads printed and handwritten text in the same request; handwriting confidence tends to be lower and should be checked.
- Tested as
- Response-reading questions ask where the line text or coordinates sit.
- Trap
- Expecting key-value pairs from Read; field extraction is Document Intelligence's job.
-
Bounding box values
Object results give a bounding box as x, y, width and height in pixels from the top-left of the image.
- Remember
- Coordinates are pixels, not percentages, in Image Analysis 4.0; Custom Vision object detection returns normalized left, top, width and height between 0 and 1.
- Tested as
- Questions ask how to draw or crop a detected object.
- Trap
- Treating Custom Vision's normalized values as pixel positions.
-
Multiclass vs. multilabel classification
Multiclass assigns exactly one tag per image; multilabel allows several tags per image.
- Remember
- If an image can contain both a cat and a dog and you need both answers, choose multilabel.
- Tested as
- Project setup questions hinge on whether categories are exclusive.
- Trap
- Choosing multiclass when images can show several categories at once.
-
Precision and recall
Precision is correct positive predictions divided by all positive predictions; recall is correct positive predictions divided by all actual positives.
- Remember
- Raising the probability threshold makes the model more selective: precision tends to rise and recall tends to fall.
- Tested as
- Scenarios ask which way to move the threshold to reduce false alarms or missed items.
- Trap
- Lowering the threshold to cut false positives, which does the opposite.
-
Mean average precision (mAP)
mAP averages the average precision of each tag and summarizes object detection quality across thresholds.
- Remember
- Object detection also uses an overlap threshold (intersection over union) to decide whether a predicted box matches a labeled box.
- Tested as
- Object detection metric questions mention mAP and overlap.
- Trap
- Looking for mAP on a classification project.
-
Training data guidance
Custom Vision needs labeled images per tag, with Microsoft recommending at least 30 per tag for a reasonable first model and requiring a minimum to train.
- Remember
- Balance tags and vary lighting, angle, background and size; imbalance skews predictions toward the larger tag.
- Tested as
- Questions about poor accuracy on one tag usually point to more or more varied images for it.
- Trap
- Retraining with the same images and expecting a different result.
-
Training vs. prediction resource
Custom Vision uses a training resource to build iterations and a prediction resource to serve a published iteration.
- Remember
- Publishing names the iteration and links it to a prediction resource; client calls use the prediction key, endpoint, project ID and published name.
- Tested as
- Ordering and code questions test the publish step and the prediction call.
- Trap
- Calling an unpublished iteration or using the training key.
-
Compact domains and export
Compact domains produce smaller models that can be exported to ONNX, TensorFlow, Core ML or a Docker container.
- Remember
- Choose a compact domain when the model must run offline on a phone, camera or edge device.
- Tested as
- Edge and offline requirements point here.
- Trap
- Training on a non-compact domain and then trying to export.
-
Video Indexer insights
Video Indexer combines speech, vision and language models to produce transcripts, OCR, labels, keywords, topics, scenes, shots, keyframes and faces.
- Remember
- Results come as an insights JSON and can be embedded through player and insights widgets; face features require limited-access approval.
- Tested as
- Searchable video archives and training-video transcripts point here.
- Trap
- Building a custom frame-by-frame pipeline with Image Analysis for something Video Indexer already does.
-
Spatial Analysis operations
Spatial Analysis runs operations such as person count, person crossing line, person crossing polygon and person distance on camera feeds.
- Remember
- It runs as an edge container on a GPU-equipped device and emits events, without identifying individuals.
- Tested as
- Occupancy, queue and distancing scenarios point to a specific operation.
- Trap
- Choosing Video Indexer for real-time occupancy counting from live cameras.
Side by side
Concepts candidates mix up, compared on the point that decides the answer.
Tags vs. objects vs. Custom Vision object detection
| Concept | Meaning | Key difference | Common trap |
|---|---|---|---|
| Tags | Keywords describing image content | No location, generic vocabulary | Using tags when position is needed |
| Objects (prebuilt) | Detected items with bounding boxes | Location included, generic labels only | Expecting brand or product names |
| Custom Vision object detection | Your own labels with bounding boxes | Trained on your images for domain-specific items | Choosing classification when location or count matters |
Image Analysis Read vs. Document Intelligence vs. Video Indexer OCR
| Concept | Meaning | Key difference | Common trap |
|---|---|---|---|
| Image Analysis Read | Text lines and words from a single image | Fast, synchronous, printed and handwritten | Expecting form fields or tables |
| Document Intelligence | Text, layout, tables and fields from documents | Multi-page PDFs and structured extraction | Using it for scene text in photos |
| Video Indexer OCR | On-screen text across video frames with timestamps | Links text to when it appears | Sampling frames manually into Image Analysis |
Precision vs. recall vs. probability threshold
| Concept | Meaning | Key difference | Common trap |
|---|---|---|---|
| Precision | Of the items predicted as a tag, how many were right | Falls with false positives | Confusing it with recall |
| Recall | Of the real items with a tag, how many were found | Falls with false negatives | Ignoring it for safety-critical detection |
| Probability threshold | Minimum confidence to count as a prediction | Higher threshold trades recall for precision | Moving it the wrong way |
Exam traps
Each trap is an answer that looks right. Learn the clue that should change your mind.
-
Classification for counting
- What candidates pick
- A Custom Vision classification project to count defects on a circuit board.
- Why it looks right
- Classification is simpler to label and train.
- Why it's wrong
- Classification labels the whole image; counting and locating items needs object detection with bounding boxes.
Clue to watch for Words such as where, how many, locate or mark.
-
Prebuilt objects for brands
- What candidates pick
- Image Analysis objects to identify the company's own product models.
- Why it looks right
- It returns bounding boxes with no training.
- Why it's wrong
- Prebuilt objects use a generic vocabulary; product-specific labels need a custom model.
Clue to watch for Product names, SKUs or company-specific parts.
-
Threshold moved the wrong way
- What candidates pick
- Lower the probability threshold to reduce false alarms.
- Why it looks right
- Changing the threshold is the right lever.
- Why it's wrong
- Lowering it accepts more low-confidence predictions, which increases false positives; raise it to improve precision.
Clue to watch for Complaints about too many incorrect detections.
-
Export from a standard domain
- What candidates pick
- Train on the General domain, then export to ONNX for an offline camera.
- Why it looks right
- General often gives the best accuracy.
- Why it's wrong
- Only compact domains can be exported; you must train on a compact domain.
Clue to watch for Offline, on-device or edge deployment of a custom model.
-
Read for invoices
- What candidates pick
- Image Analysis Read to extract invoice totals and vendor names.
- Why it looks right
- Read does extract all the text.
- Why it's wrong
- Read returns lines of text, not labeled fields; Document Intelligence's prebuilt invoice model returns named fields with confidence.
Clue to watch for Named fields such as total, date or vendor.
-
Video Indexer for live occupancy
- What candidates pick
- Video Indexer to alert when a room exceeds capacity.
- Why it looks right
- It is the video service most candidates know.
- Why it's wrong
- Occupancy alerts from live camera feeds are Spatial Analysis person count or zone operations running at the edge.
Clue to watch for Real-time counts, queues, zones or line crossings from cameras.
Answer framework
Use this sequence for any image or video question on AI-102.
- Identify the medium and timing: single image, document, recorded video or live camera feed, because each points to a different service.
- Decide whether generic labels are enough; if the items are company-specific, move from prebuilt Image Analysis to a custom model.
- Choose classification when you only need what is in the image, and object detection when you need where or how many.
- Read metrics against the business cost: raise the threshold when false positives hurt, lower it or add data when misses hurt.
- Confirm deployment details: requested features, region support for captions, prediction resource and published name, or compact domain for edge export.
Practice question
Want more like this? The AI-102 practice test has questions for every domain.
Checklist
0 of 12
Tick each one you can do without notes. Your ticks stay in this browser.
Study strategy
Start with Image Analysis in Vision Studio or the Foundry portal. Run the same photo through each visual feature and compare the outputs, then call the API from code with two features and read the JSON. Do the same with a page of handwriting to see lines, words and confidence values.
Next, spend most of your time on Custom Vision, the highest-priority topic here. Build one classification project and one object detection project with a small set of your own images. Train, open the performance tab, move the probability threshold slider and watch precision and recall change. Publish an iteration and call it from code with the prediction key. This exercise covers project types, metrics, publishing and consumption, which together generate many questions.
Commonly underestimated areas are video and edge: what Video Indexer returns, how its widgets embed insights, which Spatial Analysis operation fits a described camera scenario, and compact domains for export. Read the overview pages and write one scenario for each.
Afterwards, practice mixed questions that pit Image Analysis, Custom Vision, Document Intelligence and Video Indexer against each other, then move to Domain 6, where OCR and image analysis become enrichment skills.
FAQ
How much of AI-102 is computer vision?
Implement computer vision solutions is weighted at 10-15% in the final AI-102 study guide. Vision features also appear as built-in skills in Azure AI Search scenarios.
When do I use Image Analysis instead of Custom Vision?
Use Image Analysis when generic labels, captions, objects or text are enough. Use Custom Vision when you need your own categories, such as product models or defect types, that the prebuilt vocabulary does not cover.
How should I remember precision and recall?
Precision asks how many of the model's positive predictions were right; recall asks how many real positives the model found. Raising the threshold usually improves precision and lowers recall.
Can Image Analysis read handwriting?
Yes. The Read feature extracts printed and handwritten text from the same image. For multi-page documents, tables or named fields, use Document Intelligence instead.
What is the difference between Video Indexer and Spatial Analysis?
Video Indexer extracts searchable insights such as transcripts, text, labels and scenes from video content. Spatial Analysis watches live camera feeds at the edge to count people and detect movement through zones and lines.
Is Custom Vision still worth studying?
It is in the final AI-102 outline, so yes for this exam. Microsoft has announced Custom Vision will retire on September 25, 2028, with Azure Machine Learning AutoML and Content Understanding among the suggested alternatives.
How does this domain connect to knowledge mining?
In Azure AI Search, OCR and image analysis run as built-in skills in a skillset, so the vision features learned here decide what text and tags get indexed from images and scanned files.
Other Artificial Intelligence & Machine Learning certifications
Where the same subject shows up in other exams. Cards marked "Same subject" open the matching domain.
Test your Domain 4 knowledge
You have just compared prebuilt image features with custom models, linked precision and recall to the probability threshold, and separated Video Indexer from Spatial Analysis. Practice questions test whether you can pick the right one from a short requirement.
Keep going
- Training Training Built on your Weak areas
- Practice test Questions by domain, with explanations
- Mock exam Full length, timed like the real exam
- Readiness test Find your weakest domain in minutes
- Cheat sheet Key terms and formulas on one page
- Ask the AI-102 AI tutor Get any concept explained your way