VisAge: Age and Gender Classification from Face Images

Overview

VisAge is a comprehensive machine learning project focused on gender and age classification from facial images using the Adience Benchmark dataset. The project explores traditional machine learning and deep learning approaches, compares different feature extraction techniques, and includes a real-time camera application for age and gender prediction.

You can see the project report here.

Features

• Complete data exploration and preprocessing pipeline for the Adience dataset
• Feature extraction using ResNet50 and PCA
• Implementation of various classification models:
  • Traditional ML: Softmax, SVM, Random Forest, AdaBoost, KNN
  • Neural Networks: MLP, CNN
• Parallel processing pipelines for RGB and grayscale images
• Extensive model evaluation and performance analysis
• Research on model fairness across different demographics
• Real-time camera application for age and gender prediction

Dataset

The project uses the Adience Benchmark dataset for gender and age classification, which contains:

• ~19,000 face images of 2,284 subjects
• Age labels across 8 ranges: (0-2), (4-6), (8-12), (15-20), (25-32), (38-43), (48-53), (60+)
• Gender labels: Male, Female
• Various lighting conditions, poses, and image quality

Project Structure

VisAge/
├── data_exploration.ipynb      # Dataset analysis and preprocessing
├── img2vec.py                  # Feature extraction using ResNet50
├── grayscale_models.ipynb      # Models trained on grayscale images
├── rgb_models.ipynb            # Models trained on RGB images
├── questions/
│   └── q1.ipynb                # Analysis of model fairness across demographics
├── old_models/                 # Earlier model implementations
├── camera-viewer-app/          # Real-time age and gender classification application
│   ├── src/
│   │   ├── main.py             # Application entry point
│   │   ├── ui/                 # User interface components
│   │   └── utils/              # Utility functions
│   └── models/                 # Pre-trained models
└── docs/                      # Project report and documentation

Installation

Prerequisites

• Python 3.7+
• CUDA-compatible GPU (recommended for faster training)

Setup

1. Clone the repository:

git clone https://github.com/ShayGali/VisAge.git

2. Create a virtual environment:

python -m venv venv
# for Windows
.\venv\Scripts\activate

# for MacOS/Linux
source venv/bin/activate

3. Install dependencies:

pip install -r requirements.txt

4. Download the dataset:

• Go to Adience dataset on Kaggle
• Download and extract to the data folder
• Go to UTKFace dataset on Kaggle
• Download and extract to the data folder

Usage

Data Preprocessing

Run the data exploration notebook to preprocess the dataset:

jupyter notebook data_exploration.ipynb

Model Training

Execute the notebooks to train and evaluate models:

jupyter notebook rgb_models.ipynb
jupyter notebook grayscale_models.ipynb

Camera Application

Run the real-time age and gender classification application:

cd camera-viewer-app
python src/main.py

Model Performance

RGB Models (Best Results)

Model	Test Accuracy
MLP	34.78%
CNN	32.98%
Softmax	32.37%
Random Forest	32.18%
AdaBoost	31.62%
KNN (k=9)	23.33%
SVM	17.51%

Grayscale Models

Model	Test Accuracy
MLP	32.29%
CNN	31.53%
Random Forest	31.67%
Softmax	31.35%
AdaBoost	26.50%
KNN (k=9)	24.40%
SVM	17.81%

Research Findings

Model Performance Across Demographics

• Performance varies across racial groups, with generally lower accuracy for Black faces
• The Softmax model showed the most consistent performance across demographics

Real-Time Application Performance

• Models work efficiently in real-time with standard webcam hardware
• Performance remains stable with varying lighting conditions and when subjects wear glasses
• Processing every 5th frame maintains good balance between performance and computational efficiency

Age and Gender Classification Insights

• Gender classification is generally more accurate than age classification
• Most classification errors occur between adjacent age groups
• The MLP model demonstrated the best balance between accuracy and generalization

Camera Viewer Application

The real-time camera application detects faces and predicts age and gender using our best-performing model.

Features:

• Real-time face detection using OpenCV
• Age and gender prediction with confidence scores
• Green bounding boxes around detected faces
• Efficient processing for smooth performance on standard hardware

Challenges and Limitations

Data Challenges:

• Inconsistent age labeling in the original dataset
• Class imbalance across age groups
• Limited diversity in the dataset

Technical Challenges:

• Balancing feature complexity with model performance
• Tuning early stopping parameters to prevent overfitting
• Combined age-gender prediction complexity

Application Demo

Contributors

• Michael Dadush
• Shay Gali

License

This project is released under the MIT License.