A simple project exploring facial recognition using OpenCV and Tensorflow 2 with Keras. A special thanks to Sefik Serengil for his deeepface repo, which inspired this effort. The necessary dependencies can be installed from the repo directory using:

pip install -r requirements.txt

The system consists of the following major components:

• A face detector
• A face encoder
• A similarity search through a repository of known faces to enable recognition

OpenCV's Haar Cascade detector is used for face detection by default, with MTCNN as an alternative. The Haar Cascade is preferred for real time detection (especially if running on a laptop or personal PC), but I found MTCNN to be more robust across various lighting conditions.

This is achieved using the vgg-face model. The weights can be found in Sefik's repo of pretrained weights.

Once a face has been detected and encoded into a vector representation, it is then compared to a collection of reference faces using a linear search with a Euclidean distance threshold. The simplest setup is to create and populate ./references in this repo using a structure like the one shown below.

references
    |
    |__Person_A
    |   |__sijvbiu.jpg
    |   |__fjvbked.jpg
    |   |__bvjuhhd.jpg
    |
    |__Person_B
        |__7f84gfb.jpg
        |__cbge77d.jpg
        |__vkjffnn.jpg

In the above example 'Person_A' and 'Person_B' are the names of people of interest. This can be done programmatically using manage_reference_faces.py.

manage_reference_faces.py --person Joe_Blogs --add /path/to/folder/with/example/faces

People can also be removed.

manage_reference_faces.py --person Joe_Blogs --remove

Alternatively, references can be converted into a SQLite database of face embeddings. Note: ./references must have already been populated in order to perform this operation.

manage_reference_faces.py --build-database

This will produce the SQLite database face_embeddings.db with a single table called face_embeddings. The facial recogniser will default to using the database if it has been created.

Webcam Video Stream

main.py

Image

main.py --image /path/to/image

Image with specific Location for reference faces (this will override the use of the SQLIte database if it exists).

main.py --references /path/to/references

Server-Client configuration

Note: By default the client will try to connect to the server on 127.0.0.1:8000 unless specified as shown below.

Server

server.py --ip IP_ADDRESS --port PORT

Client with video

client.py --ip IP_ADDRESS --port PORT

Client with video and client side visualisation

client.py --ip IP_ADDRESS --port PORT --show-image

Client with image

client.py --ip IP_ADDRESS --port PORT --face /path/to/image/with/face

Client with image and client side visualisation

client.py --ip IP_ADDRESS --port PORT --face /path/to/image/with/face --show-image