Main Concept
Amazon Rekognition is a fully managed computer vision service that uses ML to analyze images and videos — identifying objects, people, text, scenes, and faces automatically. No ML expertise or model training required for standard capabilities.
Key Idea
Input → images or video files.
Output → labels, attributes, detections, and analysis about the visual content.
Underlying technology → computer vision powered by deep learning.
Core Pre-built Capabilities
Object and Scene Labeling
Detects and labels what is in an image — objects, animals, activities, environments.
Example
Image of a mountain trail: labels → Person, Rock, Outdoors, Mountain Bike, Crest.
Use case: automatically tag and categorize large media libraries without human review.
Face Detection and Analysis
Detects faces in images and extracts attributes from them.
Attributes detected: gender, age range, emotions (happy, sad, angry),
eyes open/closed, smiling, wearing glasses, and more.
Example
Use case: analyze audience photos at an event to understand demographics and emotional response.
Face Search and Verification
Compares a detected face against a database of known faces — for identity verification or people counting.
Example
Use case: verify that the person in a selfie matches the ID photo on file during onboarding.
Celebrity Recognition
Automatically identifies well-known public figures in images and videos.
Example
Rekognition detects Werner Vogels in a conference photo without any prior training.
Use case: automatically tag celebrity appearances in media archives.
Text Detection
Reads and extracts text that appears within images or video frames.
Example
Use case: detect racer numbers on vehicles in a sports broadcast to track positions.
Face Liveness
Determines whether the face shown in a video call or photo is a real, live person — not a photo, mask, or deepfake.
Example
Use case: prevent identity fraud during remote onboarding by verifying the person is physically present.
Pathing
Tracks the movement path of people or objects across video frames over time.
Example
Use case: sports analytics — track the path a player or ball takes across the field during a game.
Two Custom Capabilities
Custom Labels
Train Rekognition to recognize YOUR specific objects, logos, or products that the pre-built model does not know about. Requires only a few hundred labeled images.
Key Idea
Pre-built labeling → detects general objects (person, car, tree).
Custom Labels → detects YOUR specific objects (your logo, your products, your proprietary categories).
Process → label images → store in Amazon S3 → train Custom Labels model → deploy.
Example
The NFL uses Rekognition Custom Labels to automatically detect their logo appearing in social media images posted by fans — something the generic model cannot do since it does not know what the NFL logo looks like.
Process: upload hundreds of labeled logo images to S3 → train Custom Labels → analyze social media images → Rekognition flags every image containing the logo.
Content Moderation
Automatically detects inappropriate, unwanted, or offensive content in images and videos. Reduces the volume of content requiring human review to only 1–5% of total volume.
Key Idea
Rekognition → handles the automated moderation of 95–99% of content.
Amazon Augmented AI (A2I) → handles the human review of the remaining 1–5% that Rekognition is uncertain about.
Feedback loop → human review results from A2I can be fed back into retraining Rekognition, improving accuracy over time.
Example: Content moderation in a chatbot pipeline
User requests an image from a generative AI chatbot → chatbot generates the image → image is sent to Rekognition DetectModerationLabels API → if labels are clear of harmful content, image is returned to the user → if flagged, image is blocked or sent for human review via Amazon A2I.
Custom Moderation Adapter
Extends the default content moderation capability with your own labeled images — defining what is acceptable or not acceptable specifically for your use case or audience.
Example
A children’s education platform has stricter content standards than the default moderation model. A Custom Moderation Adapter trains Rekognition on the platform’s specific standards — flagging content that the default model would pass.
Common Exam Scenarios
Key Idea: When the answer is Amazon Rekognition
“Automatically detect inappropriate images uploaded by users” → Content Moderation.
“Verify a user’s identity by comparing their selfie to their ID photo” → Face Search and Verification.
“Detect our company logo appearing in social media posts” → Custom Labels.
“Analyze sports footage to track player movements” → Pathing.
“Prevent fake identity documents during remote onboarding” → Face Liveness.
“A moderation system needs human review for uncertain cases” → Rekognition + Amazon A2I.
Any scenario involving analyzing images or video with computer vision → Amazon Rekognition.
Rekognition + Amazon A2I Integration
Key Idea
Rekognition and Amazon Augmented AI (A2I) are designed to work together for content moderation:
Rekognition → automated decision for clear cases (95–99% of volume).
A2I → human review for uncertain cases (1–5% of volume).
Feedback loop → human decisions feed back into Rekognition training, continuously improving accuracy.
Exam Scope
You will not be asked how to implement Rekognition. You need to:
- Know what Rekognition does (computer vision — images and video).
- Recognize all core capabilities and match them to scenarios.
- Distinguish Custom Labels (recognize your objects) from Content Moderation (filter harmful content).
- Know that A2I handles the human review layer for uncertain moderation cases.
- Know the Custom Moderation Adapter extends default moderation for specific use cases.
- Know that content moderation reduces human review to 1–5% of content volume.
Exam Domain
- Domain 1, Task Statement 1.2: “Explain the capabilities of AWS managed AI/ML services (for example, Amazon Rekognition).”
- Domain 1, Task Statement 1.2: “Identify examples of real-world AI applications (for example, computer vision).”
- Domain 4, Task Statement 4.1: content moderation connects to responsible AI and safety tools.