Face Recognition

In the Getting Started , we had an overview of the face recognition API. In this section, we shall explore all the functionalities of the API.

Starting DeepStack on Docker

Run the command below as it applies to the version you have installed

docker run -e VISION-FACE=True -v localstorage:/datastore -p 80:5000 deepquestai/deepstack

Basic Parameters

-e VISION-FACE=True This enables the face recognition APIs.

-v localstorage:/datastore This specifies the local volume where DeepStack will store all data.

-p 80:5000 This makes DeepStack accessible via port 80 of the machine.

Face Registration

The face registration endpoint allows you to register pictures of person and associate it with a userid. You can specify multiple pictures per person during registration.

Example

import requests

user_image1 = open("image1.jpg","rb").read()
user_image2 = open("image2.jpg","rb").read()

response = requests.post("http://localhost:80/v1/vision/face/register",
files={"image1":user_image1,"image2":user_image2},data={"userid":"User Name"}).json()

print(response)

Response

{'message': 'face added', 'success': True}

The response above indicates the call was successful. You should always check for the ** success ** status. If there is an error in your request, you will receive a response like

{'error': 'user id not specified', 'success': False}

This indicates that you omitted the userid in your request. If you omitted the image, the response will be

{'error': 'No valid image file found', 'success': False}

Face Recognition

The face recognition endpoint detects all faces in an image and returns the userid for each face. Note that the userid was specified during the registration phase. If a new face is encountered, the userid will be unknown.

We shall test this on the image below.

../_images/idriselba2.jpg
import requests

image_data = open("test-image2.jpg","rb").read()

response = requests.post("http://localhost:80/v1/vision/face/recognize",
files={"image":image_data}).json()

for user in response["predictions"]:
   print(user["userid"])

print("Full Response: ",response)
Idris Elba
unknown
Full Response:  {'success': True, 'predictions': [{'x_min': 215, 'confidence': 0.76965684, 'x_max': 264, 'y_max': 91, 'y_min': 20, 'userid': 'Idris Elba'}, {'x_min': 115, 'confidence': 0, 'x_max': 162, 'y_max': 97, 'y_min': 31, 'userid': 'unknown'}]}

As you can see above, the first face is unknown since we did not previously register her, however, Idris Elba was detected as we registered a picture of him in the previous tutorial. Note also that the full response contains the coordinates of the faces.

Extracting Faces

The face coordinates allows you to easily extract the detected faces. Here we shall use PIL to extract the faces and save them

import requests
from PIL import Image

image_data = open("test-image2.jpg","rb").read()
image = Image.open("test-image2.jpg").convert("RGB")

response = requests.post("http://localhost:80/v1/vision/face/recognize",
files={"image":image_data}).json()

for face in response["predictions"]:

   userid = face["userid"]
   y_max = int(face["y_max"])
   y_min = int(face["y_min"])
   x_max = int(face["x_max"])
   x_min = int(face["x_min"])
   cropped = image.crop((x_min,y_min,x_max,y_max))
   cropped.save("{}.jpg".format(userid))
../_images/idriselba3.jpg
../_images/unknown.jpg

Setting Minimum Confidence

DeepStack recognizes faces by computing the similarity between the embedding of a new face and the set of embeddings of previously registered faces. By default, the minimum confidence is 0.67. The confidence ranges between 0 and 1. If the similarity for a new face falls below the min_confidence, unknown will be returned.

The min_confidence parameter allows you to increase or reduce the minimum confidence.

We lower the confidence allowed below.

import requests

image_data = open("test-image2.jpg","rb").read()

response = requests.post("http://localhost:80/v1/vision/face/recognize",
files={"image":image_data},data={"min_confidence":0.40}).json()

for user in response["predictions"]:
   print(user["userid"])

print("Full Response: ",response)
Idris Elba
Adele
Full Response:  {'success': True, 'predictions': [{'userid': 'Idris Elba', 'y_min': 154, 'x_min': 1615, 'x_max': 1983, 'confidence': 0.76965684, 'y_max': 682}, {'userid': 'Adele', 'y_min': 237, 'x_min': 869, 'x_max': 1214, 'confidence': 0.6044803, 'y_max': 732}]}

By reducing the allowed confidence, the system detects the first face as Adele. The lower the confidence, the more likely for the system to make mistakes. When the confidence level is high, mistakes are extremely rare, however, the system may return unknown always if the confidence is too high.

For security related processes such as authentication, set the min_confidence at 0.7 or higher .

Managing Registered Faces

The face recognition API allows you to retrieve and delete faces that have been previously registered with DeepStack.

Listing faces

import requests
faces = requests.post("http://localhost:80/v1/vision/face/list").json()

print(faces)

Response

{'success': True, 'faces': ['Tom Cruise', 'Adele', 'Idris Elba', 'Christina Perri']}

Deleting a face

import requests

response = requests.post("http://localhost:80/v1/vision/face/delete",
data={"userid":"Idris Elba"}).json()

print(response)

Reponse

{'success': True}

Having deleted Idris Elba from our database, we shall now attempt to recognize him in our test image.

import requests

image_data = open("test-image2.jpg","rb").read()

response = requests.post("http://localhost:80/v1/vision/face/recognize",
files={"image":image_data}).json()

for user in response["predictions"]:
   print(user["userid"])

Response

unknown
unknown

Performance

DeepStack offers three modes allowing you to tradeoff speed for performance. During startup, you can specify performance mode to be , High , Medium and Low.

The default mode is Medium.

You can specify a different mode during startup as seen below as seen below

docker run -e VISION-FACE=True -e MODE=High -v localstorage:/datastore -p 80:5000 deepquestai/deepstack

Speed Modes are not available on the Raspberry PI Version