Computer Vision on Azure for AI-103: Image Analysis 4.0, Captions, Tags, Object Detection, Custom Vision, OCR and Read API, Face API, Spatial Analysis, Content Understanding, and Choosing the Right Vision Service

Leave a Comment / AI/ML, Azure

Table of Contents

In the previous post, we built AI agents with tools and memory. Now we move to Domain 3: Computer Vision — teaching machines to see and understand images, video, and physical spaces. This domain is 10-15% of AI-103, but the concepts are testable and practical.

Analogy — Teaching a robot to see. Imagine training a robot worker for a warehouse. First, it learns to describe what it sees: “a shelf with boxes” (image captioning). Then it learns to identify specific objects: “three red boxes and one blue box” (object detection). Then it learns to read labels: “Box #4521, fragile” (OCR). Then it learns to recognize faces: “that is employee John, authorized for this area” (Face API). Then it learns to understand the space: “three people in aisle 4, one has been standing there for 10 minutes” (spatial analysis). Each capability builds on the last, and Azure AI Vision provides all of them through a unified API.

Computer Vision Through a Data Engineer’s Lens

Why data engineers need to understand computer vision:

  1. PIPELINE INTEGRATION
     Your data pipelines may need to process images:
     - Extract text from scanned invoices (OCR → structured data)
     - Tag product images for a catalog (Image Analysis → metadata)
     - Read license plates from parking lot cameras (OCR → events)
     - Extract data from forms (Document Intelligence → tables)

  2. RAG ENRICHMENT
     Images in documents need to be processed for search:
     - OCR skill in AI Search skillsets (extract text from image-heavy PDFs)
     - Image captioning for multimodal search (describe images for retrieval)
     - Content Understanding for video transcription and indexing

  3. AI-103 QUESTIONS
     10-15% of the exam tests vision capabilities
     Scenario-based: "A company needs to [task]. Which service should they use?"
     Know WHEN to use each service, not just WHAT each service does

Real-world examples data engineers encounter:
  Healthcare: extract text from handwritten prescriptions (OCR)
  Retail: count customers in stores (Spatial Analysis)
  Manufacturing: detect defective products on assembly lines (Custom Vision)
  Insurance: analyze damage photos from claims (Image Analysis)
  Finance: extract data from paper checks and invoices (OCR + Document Intelligence)
  Logistics: read shipping labels and barcodes (OCR)

Azure AI Vision — The Unified Vision Service

Azure AI Vision (formerly Computer Vision) is the unified service
for all image and video understanding capabilities.

Available through:
  - Foundry Tools (the new name for Vision within Foundry)
  - REST API and SDKs (Python, C#, JavaScript, Java)
  - AI Search skillsets (OCR, image analysis during indexing)

Core capabilities:
  Image Analysis 4.0: captions, tags, objects, people, smart crop
  Custom Vision: train your own classifiers and detectors
  OCR / Read API: extract printed and handwritten text
  Face API: detect, verify, identify, analyze faces
  Spatial Analysis: people counting, movement tracking, zone monitoring
  Content Understanding: multimodal extraction from images, video, documents

Image requirements:
  Formats: JPEG, PNG, GIF, BMP
  Max file size: 4 MB
  Min dimensions: 50 x 50 pixels
  Max dimensions: 10,000 x 10,000 pixels (Read API)

Image Analysis 4.0 — Understanding What Is in an Image

Image Analysis 4.0 is the latest version of Azure’s image understanding API. One API call returns multiple insights: captions, tags, objects, people, and more.

Analogy — A detailed police report from a witness. Imagine a witness describing a scene. Image Analysis provides: the overall scene description (caption: “a busy intersection with cars and pedestrians”), keywords (tags: cars, pedestrians, traffic light, crosswalk), specific objects with locations (object detection: “a red car at coordinates x1,y1,x2,y2”), people present (people detection: “3 people detected”), and a focus point (smart crop: “the most important area is the crosswalk”).

# Image Analysis 4.0 -- single API call, multiple features
from azure.ai.vision.imageanalysis import ImageAnalysisClient
from azure.ai.vision.imageanalysis.models import VisualFeatures
from azure.identity import DefaultAzureCredential

client = ImageAnalysisClient(
    endpoint="https://your-vision.cognitiveservices.azure.com",
    credential=DefaultAzureCredential()
)

result = client.analyze(
    image_url="https://example.com/warehouse-photo.jpg",
    visual_features=[
        VisualFeatures.CAPTION,
        VisualFeatures.DENSE_CAPTIONS,
        VisualFeatures.TAGS,
        VisualFeatures.OBJECTS,
        VisualFeatures.PEOPLE,
        VisualFeatures.SMART_CROPS,
        VisualFeatures.READ       # OCR included in the same call
    ]
)

print(f"Caption: {result.caption.text}")
# "A warehouse with shelves of boxes and a forklift"

for tag in result.tags.list:
    print(f"Tag: {tag.name} ({tag.confidence:.1%})")
# Tag: warehouse (97.2%), shelf (95.1%), box (93.8%), forklift (89.4%)

for obj in result.objects.list:
    print(f"Object: {obj.tags[0].name} at {obj.bounding_box}")
# Object: forklift at x=120, y=340, w=200, h=150

Captions and Dense Captions — Describing Images

Captions:
  Generate a single human-readable sentence describing the image
  "A group of people sitting around a conference table with laptops"
  Use for: image alt text, accessibility, search metadata

Dense Captions:
  Generate MULTIPLE captions for different REGIONS of the image
  Each caption describes a specific area with bounding box coordinates
  "A laptop on a table" (region: top-left)
  "A woman presenting at a whiteboard" (region: center)
  "A coffee cup next to a notebook" (region: bottom-right)
  Use for: detailed image understanding, region-specific search

Real-world example -- E-commerce product images:
  Upload a product photo → Image Analysis generates:
    Caption: "A black leather messenger bag with silver buckle"
    Dense captions:
      "Silver buckle clasp" (region: top-center)
      "Adjustable shoulder strap" (region: left)
      "Interior zipper pocket" (region: center)
  These captions become searchable metadata in your product catalog.

Tags — Keywords from Images

Tags extract keywords that describe the image content.
The API returns tags with confidence scores.

Example -- Construction site photo:
  Tags: construction (98%), crane (95%), building (93%), workers (91%),
        hard hat (88%), scaffolding (85%), concrete (82%), outdoor (97%)

How tags differ from captions:
  Caption: one sentence describing the overall scene
  Tags: individual keywords (nouns, adjectives) with confidence scores
  Tags are better for: search indexing, filtering, categorization

Real-world example -- Insurance claims:
  Car accident photo → Tags: car (99%), damage (95%), dent (91%),
  front bumper (88%), parking lot (86%), rain (72%)
  Pipeline: photo uploaded → tags extracted → claim auto-categorized
  by damage type → routed to appropriate adjuster

Over 10,000 recognizable concepts including:
  Objects, scenes, activities, colors, textures, food, animals, landmarks

Object Detection — Finding and Locating Objects

Object detection identifies WHAT objects are in the image AND WHERE they are.
Returns bounding box coordinates (x, y, width, height) for each object.

Caption: "A warehouse shelf"
Tags: [box, shelf, forklift, pallet]
Object detection:
  - box at (50, 100, 80, 60) -- confidence 94%
  - box at (200, 100, 85, 65) -- confidence 92%
  - forklift at (400, 250, 200, 180) -- confidence 89%
  - pallet at (100, 300, 150, 50) -- confidence 85%

The difference:
  Tags tell you "there are boxes in this image"
  Object detection tells you "there are 2 boxes, here is exactly where each one is"

Real-world example -- Retail shelf compliance:
  Camera captures shelf photo → object detection finds each product
  → compare detected products vs planogram (expected layout)
  → flag missing products or incorrect placements
  → alert sent to store manager to restock

Real-world example -- Quality control in manufacturing:
  Camera on assembly line → detect components in product photo
  → verify all 5 components present → flag if any missing
  → defective product diverted to inspection station

People Detection — Finding Humans in Images

People detection identifies humans in images with bounding boxes.
Does NOT identify WHO the person is (that is Face API).

Returns:
  - Bounding box for each person detected
  - Confidence score

Example -- Office photo:
  People detected: 4
  Person 1: (100, 50, 120, 300) -- confidence 98%
  Person 2: (350, 60, 110, 290) -- confidence 96%
  Person 3: (600, 100, 100, 250) -- confidence 94%
  Person 4: (800, 80, 115, 280) -- confidence 91%

Real-world example -- Occupancy monitoring:
  Camera in conference room → people detection counts attendees
  → data pipeline records count per 5-minute interval
  → dashboard shows room utilization over time
  → facilities team identifies underused rooms
  No facial recognition needed -- just counting people

People detection vs Face API:
  People detection: finds people (including from behind, partially visible)
  Face API: analyzes faces specifically (frontal or near-frontal required)
  Use people detection when you just need counts or positions
  Use Face API when you need facial attributes or identity

Smart Crop and Background Removal

Smart Crop:
  Automatically identifies the most important region of an image
  Generates a thumbnail focused on the key content
  Configurable aspect ratio (1:1 for social media, 16:9 for banners)

  Real-world example -- E-commerce product thumbnails:
    Upload a full product photo → smart crop generates:
      1:1 thumbnail (focused on the product, not the background)
      16:9 banner (product centered with appropriate cropping)
    No manual cropping needed for thousands of product images

Background Removal:
  Removes the background from an image, isolating the foreground subject
  Returns: alpha matte (transparency mask) or foreground image

  Real-world example -- Product catalog:
    Supplier sends product photos with random backgrounds
    → background removal isolates the product
    → clean white-background image for the website
    → automated in the data pipeline (no manual Photoshop)

Custom Vision — Training Your Own Image Models

Custom Vision lets you train your own image classification and object detection models with your own images and labels — as few as 5-15 images per class.

Analogy — Training a new quality inspector. You show the inspector 10 photos of good products and 10 photos of defective products, point out the differences, and they learn to tell them apart. Custom Vision works the same way: upload labeled examples, it trains a model, and the model can then classify new images it has never seen.

Two types of Custom Vision models:

  1. CLASSIFICATION: "What is this image?"
     Input: an image
     Output: one or more labels with confidence scores

     Example -- Manufacturing defect detection:
       Labels: "good", "scratch", "dent", "crack"
       Training: 50 images of each category
       Prediction: new photo → "scratch" (94% confidence)

     Multi-class: exactly one label per image (cat OR dog)
     Multi-label: multiple labels per image (cat AND outdoor AND sunny)

  2. OBJECT DETECTION: "What objects are here and where?"
     Input: an image
     Output: objects with labels AND bounding box coordinates

     Example -- Retail product recognition:
       Labels: "Coca-Cola", "Pepsi", "Sprite", "Water"
       Training: 50 images per product with bounding boxes drawn
       Prediction: shelf photo → "Coca-Cola" at (100,200,80,120), "Water" at (300,200,75,115)

Training process:
  1. Create a Custom Vision project (classification or object detection)
  2. Upload images (minimum 5 per class, recommended 50+)
  3. Label images (assign tags for classification, draw bounding boxes for detection)
  4. Train an iteration (minutes for small datasets)
  5. Test with new images in the portal
  6. Publish the iteration to a prediction endpoint
  7. Call the endpoint from your application

When to use Custom Vision vs Image Analysis 4.0:
  Image Analysis: recognizes 10,000+ general concepts (no training needed)
  Custom Vision: recognizes YOUR specific objects (requires YOUR training images)

  "Detect cars in photos" → Image Analysis (cars are already known)
  "Detect YOUR company's 5 product models" → Custom Vision (specific to you)
  "Classify skin conditions" → Custom Vision (domain-specific)

OCR and the Read API — Extracting Text from Images

The Read API extracts printed and handwritten text from images and documents.

Analogy — A multi-lingual speed reader. The Read API is like a human who can instantly read any document in 26+ languages, whether it is typed, handwritten, printed on a label, displayed on a screen, or scrawled on a whiteboard. It reads the text, tells you where on the page each word appears, and gives you the content in a structured format you can process in your data pipeline.

Read API capabilities:
  - Printed text in 26+ languages
  - Handwritten text (English and other languages)
  - Mixed language documents (English + French in same document)
  - Rotated, skewed, or curved text
  - Tables and structured layouts
  - Multi-page documents (up to 2,000 pages)

Output structure:
  Page → Line → Word (with bounding polygon and confidence)

Real-world example -- Invoice processing pipeline:
  1. Vendor emails PDF invoices
  2. ADF pipeline saves PDFs to Blob Storage
  3. Read API extracts text from each invoice
  4. Extracted text includes: vendor name, invoice number, line items, totals
  5. Structured data loaded to SQL database
  6. Dashboard shows invoice totals by vendor, month, status

  Without OCR: someone manually types invoice data (hours per day)
  With OCR: pipeline processes 100 invoices in minutes

Read API vs Document Intelligence:
  Read API: extracts RAW text (you parse the structure yourself)
  Document Intelligence: extracts STRUCTURED data (key-value pairs, tables)

  Use Read API: when you need the text content from any image
  Use Document Intelligence: when you need structured extraction from known document types
# OCR with Image Analysis 4.0 (synchronous, fast)
result = client.analyze(
    image_url="https://example.com/invoice.jpg",
    visual_features=[VisualFeatures.READ]
)

for block in result.read.blocks:
    for line in block.lines:
        print(f"Text: {line.text}")
        for word in line.words:
            print(f"  Word: {word.text} (confidence: {word.confidence:.1%})")

# Output:
# Text: INVOICE #4521
#   Word: INVOICE (confidence: 99.2%)
#   Word: #4521 (confidence: 98.7%)
# Text: Acme Corp
#   Word: Acme (confidence: 97.5%)
#   Word: Corp (confidence: 98.1%)
# Text: Total: $1,250.00

Face API — Detecting and Analyzing Faces

The Face API detects human faces and analyzes facial attributes. It can also verify whether two faces belong to the same person.

Face API capabilities:

  FACE DETECTION (available to all):
    Detect faces in an image with bounding boxes
    Return facial attributes: head pose, glasses, blur, occlusion
    No identification -- just "there is a face here"

  FACE VERIFICATION (restricted access):
    Compare two faces: "Are these the same person?"
    One-to-one comparison
    Use for: identity verification (compare ID photo to selfie)

  FACE IDENTIFICATION (restricted access):
    Compare a face against a group: "Who is this person?"
    One-to-many comparison
    Use for: access control, attendance tracking

  FACE GROUPING (restricted access):
    Group similar faces together from a collection
    Use for: organizing photo albums, deduplication

  FACE LIVENESS (restricted access):
    Detect if the face is a live person or a photo/video of a person
    Prevents spoofing attacks (holding up a photo of someone)
    Use for: identity verification apps, preventing fraud

IMPORTANT -- Responsible AI restrictions:
  Face IDENTIFICATION and VERIFICATION require Microsoft approval
  You must apply via the Face Recognition intake form
  Approval based on: use case, data handling, fairness considerations
  Face DETECTION (attributes, bounding boxes) is available without approval
  This is an AI-103 exam topic -- know the access restrictions

Real-world example -- Building access control:
  Employee approaches door → camera captures face
  → Face API identifies employee (approved use case)
  → Access granted → event logged → security dashboard updated
  Note: requires Microsoft approval for facial identification

Real-world example -- Identity verification for banking:
  Customer opens account → uploads ID photo + takes selfie
  → Face API verifies: "Is the selfie the same person as the ID?"
  → Liveness check: "Is this a live person, not a photo?"
  → If both pass → account creation proceeds

Spatial Analysis — Understanding Physical Spaces

Spatial Analysis uses video cameras to understand how people move through physical spaces in real time.

Analogy — An invisible store manager. Spatial Analysis is like having an invisible store manager who counts every customer entering and leaving, tracks how long people wait in line, notices when someone has been in the electronics aisle for 10 minutes (potential purchase), and alerts staff when the checkout area is overcrowded. All without personally identifying anyone.

Spatial Analysis operations:

  PEOPLE COUNTING:
    Count people entering and exiting a defined zone
    Real-time: "12 people currently in the store"
    Historical: "peak was 45 people at 2:15 PM"

  ZONE MONITORING (Person in Zone):
    Detect when people enter or exit a defined area
    "Person entered the restricted zone at 3:42 PM"
    "3 people currently in the waiting area"

  LINE CROSSING:
    Detect when people cross a virtual line
    "47 people crossed the entrance line today"
    Count directional: 30 entered, 17 exited

  DWELL TIME:
    How long a person stays in a specific area
    "Average dwell time at Display A: 2 minutes 15 seconds"
    "Person has been in Zone B for 8 minutes"

  SOCIAL DISTANCING:
    Measure distances between people
    Alert when people are too close
    (Legacy feature from COVID-19 era)

Deployment:
  Spatial Analysis runs as a Docker container on an edge device
  Processes video streams locally (no video sent to the cloud)
  Only metadata (counts, events) is sent to the cloud
  Requires: IP cameras + edge compute (Azure Stack Edge or compatible device)

Real-world examples:
  Retail: customer traffic counting, store layout optimization
  Office: meeting room occupancy, hot desk utilization
  Healthcare: patient flow monitoring, waiting room management
  Transportation: platform crowding, gate area monitoring

Content Understanding — Multimodal Extraction

Azure AI Content Understanding is a newer service that uses
generative AI to extract structured data from images, video, and documents.

Key difference from other vision services:
  Image Analysis: describes images (captions, tags, objects)
  Custom Vision: classifies images into YOUR categories
  Document Intelligence: extracts fields from known document types
  Content Understanding: extracts CUSTOM structured fields using natural language schemas

How it works:
  1. Define a schema in natural language:
     "Extract: product name, price, condition, brand from this image"
  2. Upload an image, document, or video
  3. Content Understanding returns structured JSON matching your schema

Use cases:
  - Extract product details from marketplace listings (images + text)
  - Segment video into scenes with descriptions
  - Extract custom fields from non-standard documents
  - Build RAG-ready output from video content

Content Understanding vs Document Intelligence:
  Document Intelligence: prebuilt models for known document types (invoices, receipts)
  Content Understanding: generative AI for any content type with natural language schemas
  Use Document Intelligence when a prebuilt model exists for your document type
  Use Content Understanding for custom extraction from any visual content

Choosing the Right Vision Service — Decision Guide

"I need to..." → Which service?

  Describe what is in an image
    → Image Analysis 4.0 (captions, tags)

  Find specific objects and their locations
    → Image Analysis 4.0 (object detection) for common objects
    → Custom Vision (object detection) for YOUR specific objects

  Classify images into categories
    → Custom Vision (classification) -- train with your own images

  Extract text from images or documents
    → Read API (raw text extraction)
    → Document Intelligence (structured extraction from known document types)

  Detect or identify human faces
    → Face API (detection: available to all; identification: requires approval)

  Count people or track movement in physical spaces
    → Spatial Analysis (edge container, video processing)

  Extract custom structured data from any visual content
    → Content Understanding (natural language schemas, generative AI)

  Process images during AI Search indexing
    → OCR skill (extract text from images in documents)
    → Image Analysis skill (generate tags and captions for search)

Quick reference table:
  Image Analysis 4.0:    general understanding, 10,000+ concepts, no training
  Custom Vision:          YOUR specific objects/categories, requires training
  Read API / OCR:         text extraction from any image
  Document Intelligence:  structured extraction from documents (invoices, forms)
  Face API:               face detection, verification, identification
  Spatial Analysis:       video-based people counting and movement
  Content Understanding:  generative AI multimodal extraction

Common Mistakes

  1. Using Custom Vision when Image Analysis already recognizes the object. Image Analysis 4.0 recognizes over 10,000 objects, scenes, and concepts. Training a Custom Vision model to detect “car” or “person” wastes time and money — Image Analysis does this out of the box. Use Custom Vision only for objects specific to YOUR domain that the general model does not know.

  2. Using Read API when Document Intelligence would give structured output. The Read API extracts raw text from images. If you need structured data from invoices, receipts, or forms, Document Intelligence provides prebuilt models that return key-value pairs and tables directly. You avoid writing parsing logic for standard document types.

  3. Not understanding Face API access restrictions. Face identification and verification require Microsoft approval through the Face Recognition intake form. The exam tests whether you know which Face API features are restricted and which are available to all. Detection (bounding boxes, attributes) is available to everyone. Identification and verification are not.

  4. Forgetting that Spatial Analysis runs on edge devices, not in the cloud. Spatial Analysis processes video locally on edge hardware (Azure Stack Edge or compatible device). Video streams are NOT sent to the cloud — only metadata (counts, events) is transmitted. This is important for privacy and bandwidth. The exam may ask about deployment requirements for spatial analysis.

  5. Sending images larger than 4 MB to Image Analysis. The API has a 4 MB file size limit. Production pipelines should resize or compress images before calling the API. For the Read API, documents up to 2,000 pages are supported, but each page image must be under the size limit.

  6. Not considering the confidence score when processing results. Every result (tag, object, person, word) includes a confidence score (0.0 to 1.0). Processing results without filtering by confidence means acting on low-confidence detections. Set a threshold (e.g., 0.80) and only process results above it.

  7. Using Image Analysis 4.0 for video when Spatial Analysis or Content Understanding is more appropriate. Image Analysis processes individual images, not video streams. For real-time video analysis (counting, tracking), use Spatial Analysis. For extracting structured data from video (scene segmentation, transcription), use Content Understanding.

  8. Training Custom Vision with too few or unbalanced images. A model trained with 5 images of “good product” and 50 images of “defective product” will be biased. Aim for 50+ images per class with balanced representation. Include diverse angles, lighting conditions, and backgrounds for robust models.

Interview Questions

Q: What are the main capabilities of Azure AI Vision Image Analysis 4.0? A: Image Analysis 4.0 provides captions (one-sentence image descriptions), dense captions (multiple region-specific descriptions with bounding boxes), tags (keywords with confidence scores from 10,000+ concepts), object detection (objects with bounding box locations), people detection (human detection with bounding boxes), smart crop (content-aware thumbnail generation), background removal, and synchronous OCR (text extraction). All features are accessible through a single API call, and no training is required.

Q: When should you use Custom Vision vs Image Analysis 4.0? A: Use Image Analysis 4.0 when the objects or concepts you need to detect are among the 10,000+ general concepts it already recognizes (cars, people, animals, scenes, common objects). Use Custom Vision when you need to recognize domain-specific objects or categories that the general model does not know — such as your company’s specific product models, manufacturing defect types, or custom classification categories. Custom Vision requires your own labeled training images (minimum 5 per class, recommended 50+).

Q: What is the difference between Read API and Document Intelligence? A: The Read API extracts raw text from images with word-level bounding boxes and confidence scores. It works on any image or document but returns unstructured text that you must parse yourself. Document Intelligence uses prebuilt models (invoice, receipt, ID document, W-2) to extract structured data as key-value pairs, tables, and typed fields. Use Read API when you need raw text from any image. Use Document Intelligence when you need structured extraction from known document types.

Q: What Face API features require Microsoft approval? A: Face detection (bounding boxes, facial attributes like head pose and glasses) is available to all Azure customers. Face identification (matching a face to a group), face verification (comparing two faces), face grouping, and face liveness detection require application through the Face Recognition intake form and Microsoft approval. This restriction supports Responsible AI principles and ensures facial recognition is used appropriately.

Q: How does Spatial Analysis differ from Image Analysis? A: Image Analysis processes individual static images in the cloud. Spatial Analysis processes real-time video streams on edge devices (Azure Stack Edge or compatible hardware). It counts people, monitors zones, tracks movement, and measures dwell times. Video is processed locally — only metadata (counts, events) is sent to the cloud, preserving privacy and reducing bandwidth. Use Image Analysis for individual images. Use Spatial Analysis for real-time video monitoring of physical spaces.

Q: What is Content Understanding and when would you use it instead of other vision services? A: Content Understanding uses generative AI to extract custom structured data from images, video, and documents using natural language schemas. You define what to extract in plain language (“extract: product name, price, condition from this image”), and it returns structured JSON. Use it when no prebuilt model exists for your content type, when you need custom field extraction, or when processing video (scene segmentation, RAG-ready output). Use Document Intelligence instead when a prebuilt model matches your document type (invoices, receipts).

Q: How would you build a pipeline that processes images from a data engineering perspective? A: Store incoming images in Azure Blob Storage. Create an ADF or Fabric pipeline that triggers on new files. For text extraction, call the Read API or Document Intelligence depending on document type. For image metadata, call Image Analysis 4.0 for tags and captions. Store extracted data (text, tags, metadata) in a database or Lakehouse table. For search, use AI Search with OCR and Image Analysis skills in the indexer skillset to automatically process images during indexing. Monitor the pipeline with Application Insights and set confidence score thresholds to filter low-quality results.

Wrapping Up

Computer Vision on Azure provides a spectrum of capabilities: Image Analysis 4.0 for general image understanding, Custom Vision for domain-specific classification and detection, Read API for text extraction, Face API for facial analysis (with responsible use restrictions), Spatial Analysis for real-time video monitoring, and Content Understanding for generative AI-powered multimodal extraction.

For AI-103, the key is knowing WHEN to use each service. “Extract text from scanned documents” → Read API. “Detect YOUR custom products” → Custom Vision. “Count people in a store” → Spatial Analysis. “Structured data from invoices” → Document Intelligence. “Custom fields from any visual content” → Content Understanding. The exam tests your ability to match scenarios to services, not your ability to code the API calls.

In the next post, we cover Domain 4: Text Analysis and Speech — Language Understanding, sentiment analysis, entity recognition, PII detection, translation, and speech services.

Related posts: – AI-103 Study Guide – Microsoft Foundry Platform – Security & Responsible AI – RAG Pipelines & AI Search – Data File Formats

Related Posts

Leave a Comment

Your email address will not be published. Required fields are marked *

Scroll to Top