AWS Rekognition alternatives: choose for the feature you need to replace

Google Cloud Vision is worth testing if you need image labels or text extraction. A self-hosted detector deserves a look when you need custom object detection and can maintain the deployment. Face matching requires a different shortlist: an API that finds faces may not identify people at all.

Start with the Rekognition features your application uses. Replacing one API call can be straightforward; replacing a workflow that combines text, moderation and identity checks takes a more careful comparison. This guide assesses the documented capabilities, not results from a hands-on accuracy test.

Match the replacement to the job

Imagine a marketplace that labels uploaded photos, reads text from screenshots and sends questionable images to a review queue. A detector that recognizes objects covers only part of that workflow. The replacement also needs text extraction and a moderation process, whether those come from one provider or several.

Google Cloud Vision covers several common image tasks

Cloud Vision can return labels for objects, activities and other image content. It is a sensible candidate for a searchable image library, provided its labels match how your customers search. Test that fit using your own images; the presence of a label API says little about how useful its categories will be in your product. See Google's label detection documentation.

For text, compare the required output before switching providers. Rekognition already has DetectText, which returns words and lines from images. An OCR requirement alone is therefore not a reason to leave AWS. Check whether the problem is missed text, document structure, an input limit or the cost of your existing workflow.

Google separates ordinary image text extraction from an option designed for dense documents. The latter returns page, paragraph and other structural information. That makes the OCR output format worth comparing when your application needs more than a string of recognized words.

Moderation needs its own test. Google's SafeSearch returns likelihoods for categories including adult and violent content. Decide how those categories map to your rules, then inspect both missed violations and harmless images sent for review. Copying the same numeric threshold from another service will not establish equivalent behavior.

Cloud Vision also detects faces, but does not recognize specific individuals. That distinction rules it out as a direct replacement for an enrolled-person matching workflow.

Self-hosted detection changes who owns the work

Ultralytics YOLO is a candidate when your product needs object detection and your team wants to train and run the model. Its detection documentation covers training, validation and prediction. This is a narrower choice than replacing every Rekognition feature: an object detector does not automatically provide text extraction, moderation or identity matching.

Consider a product that must find a particular kind of damaged component in customer photos. Before choosing this route, establish who will label examples, evaluate missed defects and maintain the deployed model. Include that work in the cost comparison alongside compute and storage.

Check licensing before building around it. The Ultralytics repository uses AGPL-3.0, and the vendor offers an Enterprise license for proprietary deployments. Review the terms for the code and models you intend to use; “self-hosted” does not mean unrestricted or free to operate. Start with the vendor's licensing options.

Check the maintenance history of open-source replacements

CCExtractor's Rekognition project has a broader feature list than a standalone detector. Its README lists face-related operations, text extraction, an NSFW classifier and object detection. Those are documented project capabilities, not proof of compatibility with your existing AWS calls.

The repository's latest default-branch commit was dated May 25, 2022 when checked for this guide. Its dependency file pins versions including Django 2.2.1 and TensorFlow 2.5.0. Treat it as a project to evaluate and potentially maintain yourself, not a ready-made hosted service. Before deploying it, check dependency support, model licenses, security and whether the listed endpoints still run in your environment. We have not run that deployment test.

Check access before shortlisting a face-matching service

Microsoft Face has identification and verification features, but access is restricted. Microsoft's registration policy requires eligible Microsoft-managed customers and approved use cases. These features are not available on the free tier. Confirm eligibility before spending time on a migration plan.

Keep the decision process around a match in scope, too. AWS describes CompareFaces as probabilistic and recommends human review before acting on results that affect someone's rights, privacy or access to services. A similarity score should not become automatic proof of identity.

Replacing people pathing is a separate migration

AWS discontinued Rekognition people pathing on October 31, 2025. Its migration guide describes a replacement combining YOLOv9 detection with ByteTrack tracking. This retirement applies to people pathing, not Rekognition as a whole. Separately, AWS closed Streaming Video and Bulk Image Analysis to new customers from April 30, 2026; accounts that used them in the previous 12 months keep access.

If that is the feature you used, start with the migration guide rather than a general image-API comparison. Tracking a person across video frames introduces requirements that an image-labeling endpoint does not cover. The guide is a starting implementation to evaluate, not evidence that the replacement matches your previous accuracy or operating cost.

Compare the complete bill and the cost of mistakes

A per-image price can hide differences in what is included. Google's Vision pricing bills features applied to images and counts pages in multi-page files separately. It also lists exceptions, including SafeSearch bundled with label detection. Build the estimate around the features you will call, not a single headline rate.

For a useful trial, keep the inputs and acceptance criteria fixed. Run the same representative images through the current workflow and a candidate. Include poor lighting, unusual layouts and the cases your support team already knows cause problems. Record missed results, incorrect results, response time and the amount of manual review required.

Then add the migration work: replacing API calls, mapping output fields, changing review rules and supporting the new service. Keep Rekognition if a candidate cannot show a meaningful benefit on the workload you have. A change is easier to justify when you can identify the specific failure it fixes and the work it adds.