Descripción general de la herramienta Amazon Rekognition 2026: características y precios

Resumen rápido: Amazon Rekognition is AWS’s fully managed computer vision service that uses deep learning to analyze images and videos without requiring machine learning expertise. It offers pre-trained APIs for face detection, object recognition, text extraction, content moderation, and custom label training. The service scales automatically and charges based on volume, starting at $0.00001 per image for the first million processed.

Amazon Rekognition launched in 2016 as AWS’s answer to the growing need for accessible computer vision technology. It’s designed to solve a specific problem: most developers need image and video analysis capabilities but don’t have the time, budget, or expertise to build machine learning models from scratch.

The service runs entirely in the cloud. You send images or videos through API calls, and Rekognition returns structured data about what it detected—faces, objects, text, inappropriate content, whatever you’re looking for.

Here’s the thing though—Rekognition isn’t one monolithic tool. It’s actually a suite of specialized APIs that each handle different computer vision tasks.

Core Capabilities and API Categories

Rekognition divides its features into two main categories: image analysis and video analysis. Each category contains multiple specialized APIs.

Image Analysis Features

The image analysis side handles static images. According to the official documentation, objects and faces must be at least 5% of the shorter image dimension for reliable detection. For a 1600×900 pixel image, that translates to a minimum of 45 pixels.

Label Detection identifies objects, scenes, activities, and concepts within images. It returns confidence scores for each detection, typically ranging from 0-100. Content Moderation scans for inappropriate or unwanted content across multiple categories.

Face Detection and Analysis can identify faces in images and extract attributes like estimated age range, perceived gender, emotions, and whether the person appears to be wearing glasses. Face Comparison measures similarity between two faces, while Face Search matches detected faces against a stored collection.

Text Detection extracts printed and handwritten text from images—useful for document processing, street sign recognition, and similar applications.

Video Analysis Features

Video analysis works differently. You can process stored videos asynchronously or analyze streaming video in real-time. The system detects people, tracks their movement, recognizes objects and activities, and identifies when scenes change.

For segment detection, you can filter results by confidence threshold. A 70% minimum confidence filter is commonly used to balance accuracy and coverage.

But wait. As of April 30, 2026, Amazon discontinued offering Streaming Video Analysis and Batch Image Content Moderation to new customers. Existing users who’ve accessed these features within the last 12 months retain access, but new accounts cannot enable them.

Face Liveness Detection

Face Liveness is Rekognition’s answer to spoofing attacks. The feature analyzes short selfie videos to determine whether a real person is present or if someone is using a photo, pre-recorded video, 3D mask, or deepfake.

It detects presentation attacks (printed photos, digital screens, paper masks shown to the camera) and bypass attacks (pre-recorded or synthetic videos injected into the video stream). The system returns a configurable confidence score from 0-100.

This functionality integrates with React web applications, native iOS apps, and native Android apps. No infrastructure management required—it’s fully managed by AWS.

Custom Labels for Specialized Recognition

The pre-trained models work well for general use cases. But what about specialized scenarios—detecting specific vehicle models, identifying manufacturing defects, or recognizing proprietary products?

That’s where Custom Labels come in. You provide training images labeled with your custom categories, and Rekognition builds a model tailored to your specific recognition needs. Recent improvements mean you can now build quality models with less training data than previously required.

Seven new APIs now support programmatic model creation, dataset management, and training workflow automation.

Apply Geospatial Image Analysis with FlyPix AI

Amazon Rekognition is used for image and video analysis across general computer vision tasks. FlyPix AI works in a more specific area, helping teams analyze satellite, drone, and aerial imagery to detect objects, segment locations, and monitor visible changes across real-world sites.

FlyPix AI can support location-based image analysis tasks such as:

• Detecting visible objects, roads, buildings, vehicles, vegetation, or infrastructure
• Segmenting mapped areas such as land, water, fields, and built zones
• Reviewing changes across satellite, drone, or aerial imagery over time
• Creating custom AI models for geospatial detection tasks

Contacta con FlyPix AI to discuss how geospatial image analysis can support your location-based visual review workflow.

Pricing Structure and Cost Management

Rekognition uses tiered pricing based on monthly volume. The structure changed significantly—AWS simplified tiers and reduced prices by up to 38% for high-volume users.

Group 1 APIs include most common features—AssociateFaces, CompareFaces, DisassociateFaces, IndexFaces, SearchFacesbyImage, SearchFaces, SearchUsersByImage, and SearchUsers. Group 2 APIs include DetectFaces, DetectModerationLabels, DetectLabels, DetectText, and RecognizeCelebrities with separate pricing structures.

Face vector storage costs $0.00001 per face metadata per month. Video analysis uses per-minute pricing rather than per-image.

The short answer? For moderate usage, costs stay quite manageable. 

Confidence Thresholds and Accuracy

Every detection Rekognition makes includes a confidence score. Understanding how to set appropriate thresholds matters for real-world applications.

For casual photo applications identifying family members, a threshold around 80% typically works well. High-stakes security scenarios require stricter standards—often 99% or higher to minimize false matches.

Research on facial recognition technology shows performance varies significantly based on image quality, lighting, angle, and demographic factors. Real-world performance depends on these conditions.

Research reports indicate facial recognition misidentification has contributed to wrongful arrests in documented cases, highlighting the importance of using these tools as investigative aids rather than definitive proof.

Integration and Access Control

Rekognition integrates with AWS Identity and Access Management (IAM) for access control. Policies define which users and applications can call which APIs, ensuring only authorized systems access visual analysis capabilities.

The service works through standard AWS SDKs available for Python, JavaScript, Java, .NET, and other languages. REST API access is also available for custom integrations.

You retain full ownership of content sent to Rekognition. AWS uses your data only with explicit consent and according to your configuration.

Common Use Cases

Real talk: Rekognition’s adoption spans diverse industries. Media companies use it to catalog video archives automatically. Security applications employ face search for identity verification. E-commerce platforms detect products in user-uploaded images.

Social media platforms leverage content moderation to filter inappropriate images at scale. Manufacturing operations use Custom Labels to spot defects on assembly lines. Document processing workflows extract text from scanned forms and receipts.

Washington County, Oregon, adopted Rekognition for law enforcement facial searches. According to the Washington Post, the county was paying about $7 a month for all of its searches by 2019. In 2018, the agency logged over 1,000 facial searches.

Consideraciones técnicas

Performance scales automatically—analyzing millions of images or video streams within seconds becomes routine. No infrastructure provisioning required; AWS handles capacity management.

Latency varies by operation. Simple label detection typically completes in under a second. Video analysis of stored content processes asynchronously, with results available via callback or polling.

The system works best with clear, well-lit images where subjects occupy a reasonable portion of the frame. Extreme angles, poor lighting, or heavily compressed images degrade accuracy.

Documentation and Developer Resources

AWS provides comprehensive documentation covering conceptual overviews, API references, and implementation guides. Separate developer guides exist for standard Rekognition features and Custom Labels.

Getting started typically involves creating an AWS account, configuring IAM permissions, installing an SDK, and making your first API call with a sample image. The documentation includes example code for common scenarios.

Preguntas frecuentes