Skip to main content

Notes: Recognition

Updated Jul 29, 2020 ·

AWS Rekognition​

AWS Rekognition is a computer vision service. It allows you to analyze images without building your own model.

  • Works with familiar boto3 patterns
  • Can detect objects like bicycles or cars
  • Can extract text from signs or documents

If you create your own model, you would need training data and maintenance.

  • It would only work for a specific task
  • Rekognition is general-purpose and ready to use
  • This makes it easy for data engineers to get results quickly.

Object Detection​

As example, we'll use the image below:

Before detection, upload your image to an S3 bucket.

import boto3
s3 = boto3.client(
    's3',
    region_name='us-east-1',
    aws_secret_key_id=AWS_KEYID,
    aws_secret_access_key=AWS_SECRET
    )

s3.upload_file(
    Filename='bike.jpg',
    Bucket='city-images',
    Key='bike.jpg'
)

The next step is to construct the Boto3 Rekognition client. You can use the Rekognition attributes to detect objects like bicycles.

rekog = boto3.client(
    'rekognition',
    region_name='us-east-1',
    aws_secret_key_id=AWS_KEYID,
    aws_secret_access_key=AWS_SECRET
    )

response = rekog.detect_labels(
    Image={'S3Object': {
        'Bucket': 'city-images', 
        'Name': 'bike.jpg'
        }},
    MaxLabels=10,
    MinConfidence=90
)

Optionally, we can specify the maximum amount of labels to return and the minimum confidence in the match that we are willing to accept. Rekognition returns a list of labels with instances and confidence scores.

In this image, Rekognition sees 1 bicycle and many cars.

Text Detection​

Text detection works in the same way...Detect text in images, such as signs or documents.

text_response = rekog.detect_text(
    Image={'S3Object': {
        'Bucket': 'city-images', 
        'Name': 'sign.jpg'
        }}
)

The response returns a list of dictionaries. Detections can be of 2 types:

  • Line detections combine adjacent words

  • Word detections show individual words.

AWS Translate​

AWS Translate lets you convert text between languages easily.

Using Boto3:

  • Initialize a boto3 client for Translate
  • Call translate_text with the text and target language
  • Use 'auto' for SourceLanguageCode to detect the source language automatically
import boto3

translate = boto3.client('translate')

response = translate.translate_text(
    Text="Hello, learner!",
    SourceLanguageCode='auto',
    TargetLanguageCode='es'
)

print(response['TranslatedText'])

AWS returns the translated text, detected source language, and target language.

Hola aprendiz!

Detecting Language​

You can identify the language of a text using AWS Comprehend.

  • Initialize a boto3 client for Comprehend
  • Call detect_dominant_language with the text
  • Response includes languages with confidence scores
comprehend = boto3.client('comprehend')

response = comprehend.detect_dominant_language(Text="Hola aprendiz")
print(response['Languages'])

Expected output:

## "es" refers to Spanish 
[{'LanguageCode': 'es', 'Score': 0.99}]

Detecting Sentiment​

AWS Comprehend can also analyze the sentiment of text.

  • Call detect_sentiment with your text
  • Response includes Sentiment key and confidence scores for Positive, Negative, Neutral, and Mixed
response = comprehend.detect_sentiment(Text="I love this product!", LanguageCode='en')
print(response['Sentiment'])

Expected output:

Positive

Sentiment detection helps understand user feedback or social media posts at scale.