|
--- |
|
license: mit |
|
language: |
|
- en |
|
tags: |
|
- image-classification |
|
- computer-vision |
|
- deep-learning |
|
- face-detection |
|
--- |
|
--- |
|
language: en |
|
tags: |
|
- image-classification |
|
- computer-vision |
|
- deep-learning |
|
- face-detection |
|
- resnet |
|
datasets: |
|
- custom |
|
license: mit |
|
--- |
|
|
|
# ResNet-based Face Classification Model 🎭 |
|
|
|
This model is trained to distinguish between real human faces and AI-generated faces using a ResNet-based architecture. |
|
|
|
## Model Description 📝 |
|
|
|
### Model Architecture |
|
- Deep CNN with residual connections based on ResNet architecture |
|
- Input shape: (224, 224, 3) |
|
- Multiple residual blocks with increasing filter sizes [64, 128, 256, 512] |
|
- Global average pooling |
|
- Dense layers with dropout for classification |
|
- Binary output with sigmoid activation |
|
|
|
### Task |
|
Binary classification to determine if a face image is real (human) or AI-generated. |
|
|
|
### Framework and Training |
|
- Framework: TensorFlow |
|
- Training Device: GPU |
|
- Training Dataset: Custom dataset of real and AI-generated faces |
|
- Validation Metrics: |
|
- Accuracy: 52.45% |
|
- Loss: 0.7246 |
|
|
|
## Intended Use 🎯 |
|
|
|
### Primary Intended Uses |
|
- Research in deepfake detection |
|
- Educational purposes in deep learning |
|
- Face authentication systems |
|
|
|
### Out-of-Scope Uses |
|
- Production-level face verification |
|
- Legal or forensic applications |
|
- Stand-alone security systems |
|
|
|
## Training Procedure 🔄 |
|
|
|
### Training Details |
|
```python |
|
optimizer = Adam(learning_rate=0.0001) |
|
loss = 'binary_crossentropy' |
|
metrics = ['accuracy'] |
|
``` |
|
|
|
### Training Hyperparameters |
|
- Learning rate: 0.0001 |
|
- Batch size: 32 |
|
- Dropout rate: 0.5 |
|
- Architecture: |
|
- Initial conv: 64 filters, 7x7 |
|
- Residual blocks: [64, 128, 256, 512] filters |
|
- Dense layer: 256 units |
|
|
|
## Evaluation Results 📊 |
|
|
|
### Performance Metrics |
|
- Validation Accuracy: 52.45% |
|
- Validation Loss: 0.7246 |
|
|
|
### Limitations |
|
- Performance is only slightly better than random chance |
|
- May struggle with high-quality AI-generated images |
|
- Limited testing on diverse face datasets |
|
|
|
## Usage 💻 |
|
|
|
```python |
|
from tensorflow.keras.models import load_model |
|
import cv2 |
|
import numpy as np |
|
|
|
# Load the model |
|
model = load_model('face_classification_model1') |
|
|
|
# Preprocess image |
|
def preprocess_image(image_path): |
|
img = cv2.imread(image_path) |
|
img = cv2.resize(img, (224, 224)) |
|
img = img / 255.0 |
|
return np.expand_dims(img, axis=0) |
|
|
|
# Make prediction |
|
image = preprocess_image('face_image.jpg') |
|
prediction = model.predict(image) |
|
is_real = prediction[0][0] > 0.5 |
|
``` |
|
|
|
|
|
## Ethical Considerations 🤝 |
|
|
|
This model is designed for research and educational purposes only. Users should: |
|
- Obtain proper consent when processing personal face images |
|
- Be aware of potential biases in face detection systems |
|
- Consider privacy implications when using face analysis tools |
|
- Not use this model for surveillance or harmful applications |
|
|
|
## Technical Limitations ⚠️ |
|
|
|
1. Current performance limitations: |
|
- Accuracy only slightly above random chance |
|
- May require ensemble methods for better results |
|
- Limited testing on diverse datasets |
|
|
|
2. Recommended improvements: |
|
- Extended training with larger datasets |
|
- Implementation of data augmentation |
|
- Hyperparameter optimization |
|
- Transfer learning from pre-trained models |
|
|
|
## Citation 📚 |
|
|
|
```bibtex |
|
@software{face_classification_model1, |
|
author = {Your Name}, |
|
title = {Face Classification Model using ResNet Architecture}, |
|
year = {2024}, |
|
publisher = {HuggingFace}, |
|
url = {https://huggingface.co/arsath-sm/face_classification_model1} |
|
} |
|
``` |
|
|
|
## Contributors 👥 |
|
- Arsath S.M |
|
- Faahith K.R.M |
|
- Arafath M.S.M |
|
|
|
University of Jaffna |
|
|
|
## License 📄 |
|
This model is licensed under the MIT License. |