This is a DEMO system that uses Django, PostgreSQL, OpenCV and Dlib to perform facial recognition.
The dataset used for the tests is the LFW
- Django - Python web framework
- OpenCV - Computer Vision library
- Dlib - Modern C++ toolkit containing machine learning algorithms
- Postsgresql - (with CUBE extension) - Open source object-relational database
The training mode
For the "training" of a person's face, a photo is uploaded to the system. If there is a face in the image, 128 points (features) are extracted from that face. This array of points are then stored in the database.
Facial recognition
To recognize a person's face, we pass a photograph to the system. The system will extract a matrix of 128 points from the face and then it will try to locate in the database which previously trained face has the matrix of points closest to the searched face. To do this search in the PostgreSQL database the CUBE extension is used. This extension provides a function that calculates the Euclidean distance of the face points with the points saved in the database. The query returns the id of the face with the closest array of points to the points of the queried face.
To run this demo it is necessary GIT and DOCKER / DOCKER COMPOSE installed on your machine.
If you don't have docker, just follow the step by step according to your operating system at this LINK.
Clone repository
git clone https://github.com/psilva1982/face_api.git
Starting the system
cd face_api
docker compose -f 'docker/dev/docker-compose.yml' up
-
The container image creation step takes some time. (~5 minutes 😴)
-
If you wish, it is not necessary to load and train images from the LFW dataset. The training is a very slow process, there are more than 13000 images to extract the points matrix and save it in the database. On my personal computer, this process takes an average of 90 minutes , so if you choose to go this route while the images are being processed, you can walk away a bit, grab a coffee ☕ , etc... On the other hand, if you choose not to import the dataset there is an option in django-admin to register and train your personal images for testing.
Download, extract and loading the LFW dataset into database.
docker compose -f 'docker/dev/docker-compose.yml' run --rm django bash -c "python src/manage.py load_lfw"
Training / extract points from images (more time needed... 💤)
docker compose -f 'docker/dev/docker-compose.yml' run --rm django bash -c "python src/manage.py training_photos"
Test system on browser in http://localhost:8000/ with the credentials below.
user: demo@demo.com
password: demo
Adding new persons
For add a new person go to persons list in http://localhost:8000/admin/recognizer/person/ and click add person button on top right corner.
Training new persons
Select the person you want to train the images from the persons list in http://localhost:8000/admin/recognizer/person/ and choose the action "Training person(s) photo(s)" and then click the go button.
Test facial recognition
On persons list http://localhost:8000/admin/recognizer/person/ click in facial recognition button on top right corner
See the result
😎 Enjoy !!!
- Facial recognition is more effective with more trained photos of the person you are looking for. This helps to avoid false positives. Some people in the LFW dataset only have a single photo.
- Images with more than one face cannot be trained on this system.
- To register a new user or recover password, you can access this link http://localhost:8025 (mailhog) to verify the emails received.
This project is for educational purposes only.