Arkinu
AWS service comparison

Amazon Rekognition vs Amazon Textract: differences and when to use each

Amazon Rekognition
Amazon Textract

Often confused on the examQuestions tend to offer one as a distractor for the other: look for the phrase in the scenario that tips the balance.

Machine Learning

Amazon Rekognition

On the radar· Rekognition: streaming video and bulk analysis closed to new customers

Image and video analysis (computer vision) service that requires no machine learning expertise. It detects objects, scenes, text, and inappropriate content, and analyzes and compares faces.

Exam phrases that point here

  • «identify objects, scenes, or activities in images and videos»

    Amazon Rekognition analyzes images and videos with pre-trained models: it returns labels (for example, "Beach" or "Car") with their confidence and, for common objects, their location.

    When not: If what you need to read is the text of a scanned document or a form, the answer is Amazon Textract: Rekognition text detection is designed for real-world images, such as signs or license plates.

  • «detect inappropriate content in images or videos uploaded by users»

    Rekognition content moderation flags explicit, violent, or hateful content with hierarchical labels and a confidence score, so you can filter automatically or send it to human review.

    When not: If the content to moderate is text, Rekognition does not apply: text analysis is the domain of Amazon Comprehend. And to improve detection with the business's own images, you train a Custom Moderation adapter.

  • «compare the face in a selfie with the photo on an ID document»

    CompareFaces measures the likelihood that two faces belong to the same person, the basis of an identity verification.

    When not: If you need to find a person among many stored faces, use a face collection with search (SearchFacesByImage). If you need to read the data on the ID document, Amazon Textract does that.

  • «recognize objects specific to the business, such as the brand logo or factory parts»

    Rekognition Custom Labels trains a detector for your own objects and scenes with a few labeled images, without machine learning expertise.

    When not: Custom Labels is not suitable for faces or text. If the model needs its own algorithm or data that is not images, the answer is Amazon SageMaker AI.

  • «check whether workers are wearing hard hats, face masks, or gloves»

    Rekognition personal protective equipment (PPE) detection locates face, hand, and head covers on the people in an image.

    When not: It does not identify the people or guarantee that the equipment is worn properly; for other types of equipment, such as reflective vests, you must use Custom Labels.

Often paired with

  • Amazon S3
  • AWS Lambda
Study Amazon Rekognition →

Machine Learning

Amazon Textract

Service that detects and extracts printed and handwritten text from documents, as well as forms and tables with their structure. It is used to automate data capture from documents such as invoices, receipts, or forms.

Exam phrases that point here

  • «extract text, key-value pairs, and tables from scanned documents»

    Amazon Textract goes beyond OCR: in addition to printed or handwritten text, it understands the structure of the document and returns the fields of a form and the cells of a table.

    When not: If the text appears in a real-world photo (a sign, a license plate), Amazon Rekognition text detection is the fit. If the text is already digital and what you need is to understand its meaning or sentiment, think of Amazon Comprehend.

  • «process invoices and receipts automatically»

    The Textract Analyze Expense API specializes in invoices and receipts and extracts their data without templates.

    When not: For identity documents there is Analyze ID (according to the FAQ, passports and U.S. driver's licenses). If you only need to compare the photo on the document with a selfie, that is Amazon Rekognition.

  • «multi-page PDF documents»

    Multi-page PDFs and TIFFs are processed with the Textract asynchronous operations, which read the document from S3 and notify through Amazon SNS when they finish.

    When not: For an image or a single page with an immediate response, the synchronous API is enough (for example, AnalyzeDocument), which also accepts the file's bytes and not only S3 objects.

  • «ask in natural language for a specific value in the document»

    The Analyze Document Queries feature lets you ask for a value with a question, such as "What is the due date?", without depending on how the document is laid out. According to the FAQ, Queries only processes English.

    When not: Queries returns specific data from the document; if you need to classify the text or analyze its sentiment, you then move on to Amazon Comprehend.

  • «avoid throttling when processing thousands of documents»

    The FAQ recommends retries with exponential backoff and jitter, and smoothing the traffic with an Amazon SQS queue in front of Textract.

    When not: If the sustained load exceeds the service quota, the queue is not enough: you must request a quota increase in Service Quotas.

Often paired with

  • Amazon S3
  • Amazon Comprehend
Study Amazon Textract →