Facial Demographics Analysis 👤

A robust Deep Learning application capable of detecting Age, Gender, and Race from images, videos, and live webcam feeds. This project utilizes a Multi-Task Learning approach with a shared EfficientNet backbone to perform simultaneous demographic classifications.

🚀 Features

• Multi-Task Learning: Single model architecture predicting Age, Gender, and Race simultaneously.
• Real-Time Inference: Optimized pipeline for live webcam analysis using lightweight face detection.
• Video Processing: Full video file analysis with frame-by-frame annotation and export capabilities.
• MLOps Pipeline: Modular code structure handling Data Ingestion, Preparation, and Model Training.
• Smart Face Detection: Integrates MTCNN for accurate face localization before classification.

🛠️ Tech Stack

• Frameworks: PyTorch, TensorFlow (for MTCNN), Transformers (Hugging Face)
• Base Models: EfficientNet / EfficientFormer
• Interface: Streamlit
• Data Processing: Pandas, NumPy, OpenCV, Pillow
• Dataset: FairFace (sourced via Hugging Face Hub)

📂 Project Structure

├── config/              # Configuration files (YAML)
├── src/
│   └── cnnClassifier/
│       ├── components/  # Logic for Ingestion, Prep, and Training
│       ├── pipeline/    # Orchestration of components
│       └── entity/      # Data classes for configuration
├── app.py               # Streamlit Frontend application
├── main.py              # Training pipeline entry point
├── requirements.txt     # Dependencies
└── template.py          # Project scaffolding script

⚙️ Installation

1. Clone the repository

   git clone https://github.com/AlyyanAhmed21/Facial-Age-Detection.git
   cd Facial-Age-Detection

2. Create a Virtual Environment

   conda create -n face-env python=3.8 -y
   conda activate face-env

3. Install Dependencies

   pip install -r requirements.txt

🏃‍♂️ Usage

1. Run the Web Application

To start the interface for inference (Image/Video/Webcam):

streamlit run app.py

2. Train the Model

To re-run the training pipeline (Ingestion -> Preparation -> Training):

python main.py

🧠 Model Architecture

The model utilizes a Shared Backbone strategy:

1. Input: Preprocessed Face Crop ($224 \times 224$).
2. Backbone: EfficientNet (pretrained) acts as the feature extractor.
3. Heads: Three separate Fully Connected layers branch off the pooled features:
   • Age Head: Predicts age ranges.
   • Gender Head: Predicts Male/Female.
   • Race Head: Predicts ethnicity.
4. Loss Calculation: A weighted sum loss function allows the model to learn all three tasks while prioritizing Age accuracy.

📊 Dataset

The project is trained on the FairFace dataset, which is designed to reduce racial bias in facial analysis algorithms. The pipeline automatically downloads and formats this data from the Hugging Face Hub.