Amazon Rekognition Overview of filter function and face limit when registering faces

Amazon Rekognition IndexFaces API Maximum number of faces you want to register at the time of action, It is possible to eliminate faces that are not clear but blurry.

Target person

I want to maintain quality with Amazon Rekognition
I want to save money by eliminating low quality images

With the default settings, blurred images and face images that are mixed in with the background are also subject to registration, which reduces accuracy. Traditionally, Index Faces indexes up to 15 faces in the input image. The new version of the face detection model indexes up to 100 faces in the input image.

List of filters that can be used

MaxFaces You can set the maximum value of the face image you want to register As for the registered images, clear and clear images have the highest priority.

QualityFilter Filter to exclude poor quality images If you set the QualityFilter to automatic, Amazon will exclude poor quality images.

Filter system control method

Controllable level to filter "** AUTO " if you want the filter to be automatic on the Amazon side Judgment sweet filter " LOW " Intermediate filter " MEDIUM " Strict filter " HIGH " If you do not want to filter, specify " N ONE **"

Filter usage example

`amareko_filter.py`



import boto3

collection = "qrestia"
image_file = "TestImage.jpg "
external_image_id = "Qiita test ID"

def index_faces():
    rekognition_client = boto3.client('rekognition')
    with open(image_file, 'rb') as image:
        
        rekognition_response = rekognition_client.index_faces(
            Image={'Bytes': image.read()},
            CollectionId=collection,
            ExternalImageId=external_image_id,
            QualityFilter="AUTO", #Set Quality Filter to "Automatic"
            MaxFaces=3) #Increase the maximum number of face detection to 3
    print(rekognition_response)

if __name__ == '__main__':
    index_faces()

How to check the cause of the bounce

Information about faces found in the image but not indexed can be found in the array of UnindexedFace objects. From this response, you can understand the cause and make appropriate filter adjustments.

EXTREME_POSE-The pose is such that the face cannot be detected. For example, my head is too far from the camera.
EXCEEDS_MAX_FACES-The number of detected faces is greater than the number specified by MaxFaces.
LOW_BRIGHTNESS-The image is too dark.
LOW_SHARPNESS-The image is blurry.
LOW_CONFIDENCE-Face detected with low reliability.
SMALL_BOUNDING_BOX-The border box around the face is too small.

[PYTHON] Amazon Rekognition Filter function when registering faces ・ Limiting the number of faces