This is a deep learning model to recognize faces and match them to a set of predefined faces using a siamese network.
The model uses a siamese network architecture, where we don't directly compare images but rather a relatively small representative vector computed from each image. This reduces complexity of the model, as it has to be run only once for each face to be recognized, and then simple mathematical operations can be used to do the comparison. This architecture also eliminates the need for retraining the model for new faces like classifier networks.
To train the model, populate the dataset directory with this dataset or a similarly structured one.
To start the training, run:
python3 train.py
When the training is done, you'll want to generate tflite file of the model:
python3 to_tflite.py
Populate the known_faces directory with some images of faces.
And finally to run the script to identify faces from the webcam:
python3 facial_recognition.py
The model uses is a standard ResNet50V2, with a few dense layers to compute the final vector value.
For our target use case, rejecting a known face is less undesirable than falsely identifying it. So we weighted the loss function to minimize false positives as much as possible.
| Overall accuracy | Sensitivity | Specificity |
|---|---|---|
| 79.92% | 60.47% | 99.37% |
Here, you can see a screenshot of the running model.
This project is licensed under the GPLv3 License - see the LICENSE.md file for details