Cameras are getting smarter, they're changing how we manufacture goods, how we drive our cars and even, for good or bad, how surveillance is being carried out.

And now it's getting even easier than ever to add this functionality to your solutions; it's becoming more democratised, allowing anyone to add features like facial analysis to their projects. One of the big players driving this is AWS, their goal is to put machine learning into the hands of every developer. For vision, they're doing it with Amazon Rekognition.

Amazon Rekognition

Rekognition makes it easy to add image and video analysis to your applications using pre-trained models that require no machine learning know-how. It unlocks the ability to identify objects, people, text, and actions in images and videos, as well as detecting inappropriate content.

One of the most impressive features is the highly accurate facial detection; you can use it to detect, analyze, and compare faces for a whole range of verification, behaviour and public safety use cases. This post will focus on those facial detection features and I'll cover some of the others in the future.

So, let's get started!

Pre-requisites

To try out Rekogniton you'll need to an AWS account, the team at AWS have a great guide on getting your account set-up properly with roles for developers; if you're starting fresh, you'll also need to install and configure the CLI - it'll handle storing credentials for the SDKs to use with Rekognion. For experimenting with Rekognition the full access role from the docs will work well, but for anything further, stick to granting the least possible permissions.

The examples covered in this post should fall under the AWS Free Tier but it's worth checking out the pricing page for any service before getting started.

Set-up

We'll be using Python3 so we'll need the Python AWS SDK, called boto3, to access Rekognition as well as any other AWS services,

First, you need to install boto3 using PIP;

pip install boto3

Then we can import boto into a new python file and set up a client, which acts as an interface to AWS to start using the facial detection functionality;

import boto3

# Create a Rekognition client
client=boto3.client('rekognition')

Detection

Once you've got a client set up you can call the detect_faces method with an image, the response will contain a full batch of info on any faces it detects. There's two ways to load images to Rekognition: byte arrays and S3.

S3

The way to store blobs on AWS - to use with Rekognition we just specify the bucket and file, pass it to Rekognition through our client and get our response.

response = client.detect_faces(
    Image={
         'S3Object':{
             'Bucket':bucket-name,
             'Name':photo-file-name
          }
     }
)

Using S3 allows easier integration to other services and images can be uploaded to S3 anyway you normally would.

Byte Array

We can also send images straight to Rekognition as a byte array. You can open the files as you would load any file in Python, using the built-ins;

client=boto3.client('rekognition')

with open("photo.jpg", 'rb') as image:
    response = client.detect_faces(Image={
        'Bytes': image.read()
     })

You can also try using webcams and directly firing images to Rekognition, you can access them with OpenCV, this Gist shows how you can use your Webcam with Rekognition;

Attribute Option

There's another argument you can add to the request; Attributes, you have two options ALL or Default. Default returns BoundingBox, Confidence, Pose, Quality, and Landmarks for any faces in the image, so it just detects the faces. All, takes a fraction longer and returns a fuller set of information for any faces, everything from Default and more interesting aspects like predicted emotions, age and gender.

response = client.detect_faces(
    Image={
        'S3Object':{
            'Bucket':bucket-name,
            'Name':photo-file-name
        }
    },
    Attributes=['ALL']
)

Response

After making the request, the response contains a set of "face details" for each face it detects in the image (up to around 40 faces). Here's a sample of an ALL response;

{
   "FaceDetails": [ 
      { 
         "AgeRange": { 
            "High": number,
            "Low": number
         },
         "Beard": { 
            "Confidence": number,
            "Value": boolean
         },
         "BoundingBox": { 
            "Height": number,
            "Left": number,
            "Top": number,
            "Width": number
         },
         "Confidence": number,
         "Emotions": [ 
            { 
               "Confidence": number,
               "Type": "string"
            }
         ],
         "Eyeglasses": { 
            "Confidence": number,
            "Value": boolean
         },
         "EyesOpen": { 
            "Confidence": number,
            "Value": boolean
         },
         "Gender": { 
            "Confidence": number,
            "Value": "string"
         },
         "Landmarks": [ 
            { 
               "Type": "string",
               "X": number,
               "Y": number
            }
         ],
         "MouthOpen": { 
            "Confidence": number,
            "Value": boolean
         },
         "Mustache": { 
            "Confidence": number,
            "Value": boolean
         },
         "Pose": { 
            "Pitch": number,
            "Roll": number,
            "Yaw": number
         },
         "Quality": { 
            "Brightness": number,
            "Sharpness": number
         },
         "Smile": { 
            "Confidence": number,
            "Value": boolean
         },
         "Sunglasses": { 
            "Confidence": number,
            "Value": boolean
         }
      }
   ],
   "OrientationCorrection": "string"
}

A full breakdown of the response can be found in the docs but let's take a look at some of the most interesting aspects like landmarks and emotions.

Landmarks

Landmarks are coordinates for points on the face, they're used for detecting the face itself and are points like different spots on the nose, eyes and jawline. This image from the docs shows roughly where they all are on the face:

From AWS docs

Landmarks can provide a different way of understanding detected faces, to spark some ideas of interesting ways of using landmarks; here's a rough demo to work out what a male's face shape might be;

Face Details

The details returned with the ALL attributes call can be really useful, these are aspects like estimated age range, predicted gender, if they've facial hair and if the person is wearing glasses.

Rekognition returns a dict that contains a boolean for the aspect and the confidence in the analysis. Here's a snippet for parsing the response to show the dict for detecting glasses on the faces in the scene:

for faceDetail in response['FaceDetails']:
    print(faceDetail['Eyeglasses'])

# Example Output
# {'Value': False, 'Confidence': 98.8663558959961}

Detecting aspects like glasses or age allows us to do better understand demographics to tailor services during testing or verify information about who our users could be!

Emotions

One of the most interesting aspects of Rekognition is the emotional analysis it can do. Understanding your user with the other details is useful but being able to see and understand exactly how they're reacting can create open up some really interesting opportunities.

This snippet parses the emotions of the response and displays them all, confidence indicates how likely the face is displaying that emotion:

for faceDetail in response['FaceDetails']:
    print('Emotions: \t Confidence\n')
    for emotion in faceDetail['Emotions']:

    print(str(emotion['Type']) + '\t\t' + str(emotion['Confidence']))
    print('\n')

# Sample Output
# Emotions:     Confidence
# HAPPY     0.051588401198387146
# ANGRY     7.0355730056762695
# SAD       0.04484181851148605
# DISGUSTED 0.11442316323518753
# CONFUSED  0.2994600534439087
# FEAR      0.013996988534927368
# CALM      90.5924301147461
# SURPRISED 1.8476886749267578

This on can be quite fun to play around with trying to see what gets a 'Disgusted' response rather than 'Calm' but it can be really useful, being able to even get a rough sentiment read on someone interacting with your application can be really valuable.

Streaming, Comparing and Videos

Rekognition features are a lot broader than what's been covered here but when it comes to faces, you can stream pictures in with kinesis, pass batches of images or create collections of images to sample and analyse.
You can use compare facesamong these data collections, check if the faces are celebrities, or even begin to track movement across images and video.

To take any of these further you'll want to take a proper look at the docs!

Hopefully, this has shown you that when used responsibly it can be a fantastic tool and how to start experimenting with facial detection and analysis!

Want to see some really impressive Rekognition example projects to check out? There's a bunch at the official AWS samples repo.

If you want to learn more about AWS in general? A great place to start is their Certs here's a post I wrote after getting my first!

A Journey Through the Amazon

Marc Nevin ・ Mar 2 '20

#aws #beginners #career #devops

And don't forget to take a look at other posts with the AWS Tag: