This repository is the practical part of my Final Degree Project in Computer Engineering. It hosts an exploration of the field of facial recognition using deep learning. Inspired by a crossover interest in machine learning and the visual world, this project investigates the complexities of facial recognition, focusing on model performance, speed, and efficiency in real-time implementation.
We have compared between two types of loss functions and different models of the Convolutional Neural Network.
models = [
"ResNet50",
"MobileNetV3",
"InceptionV3",
"EfficientNet",
"VGG16"
]
loss_functions = [
"Triplet Loss",
"Categorical Cross Entropy",
]
face_detector_alignment_crop = ["MTCNN"]
face_detector_real_time = ["Media-pipe"]| Base Model | Accuracy | Precision | Recall | F1-Score |
|---|---|---|---|---|
| ResNet50 - Triplet Loss | 97,78% | 97,62% | 97,78% | 97,62% |
| ResNet50 - Categorical Cross Entropy | 80% | 82,86% | 80,88% | 80,02% |
| MobileNetV3 - Triplet Loss | 91,11% | 92,98% | 92,16% | 91,63% |
| MobileNetV3 - Categorical Cross Entropy | 75,55% | 77,73% | 74,45% | 74,74% |
| InceptionV3 - Triplet Loss | 64,44% | 64,01% | 66,19% | 64,41% |
| InceptionV3 - Categorical Cross Entropy | 66,66% | 70,36% | 66,00% | 66,97% |
| EfficientNet - Triplet Loss | 95,55% | 95,56% | 95,82% | 95,46% |
| VGG16 - Categorical Cross Entropy | 84,44% | 84,44% | 84,81% | 84,53% |
In the files triplet_loss_val.ipynb and cross_entropy_val.ipynb we can visualize the results obtained with the test dataset
- Prediction on test dataset - ResNet50 - Triplet Loss
- Confusion Matrix - ResNet50 - Triplet Loss
To test models in real time we change the model path based on the missing function that we want to use in the real_time.ipynb file.
Below we have a sample of a recognition frame in real time with the model that gave us the best result (ResNet50 with the use of Triplet Loss):
- DeepFace: https://github.com/serengil/deepface
- Keras-FaceNet: https://github.com/nyoki-mtl/keras-facenet
- Face Recognition: Siamese w/ Triplet loss: https://www.kaggle.com/code/stoicstatic/face-recognition-siamese-w-triplet-loss
- Facenet: https://arxiv.org/abs/1503.03832