AgeClassification

A PyTorch-based age classification system that categorizes faces into three age groups: Under 16, 16-25, and Over 25. The project focuses on two approaches: a custom CNN model and a transfer learning model using MobileNetV2, both trained on the UTKFace dataset.

Features

• Three-Class Age Classification: Categorizes faces into age groups (Under 16, 16-25, Over 25)
• Two Model Architectures:
  • Simple CNN: Lightweight custom convolutional neural network
  • Transfer Learning: Fine-tuned MobileNetV2 for improved accuracy
• Real-time Inference: Live webcam support for age classification
• Custom Metrics: Tracks "illegal sales" (minors classified as 25+) and "customer annoyance" (adults flagged as under 25)
• Cross-platform Support: Compatible with Windows, macOS, and Linux
• Hardware Acceleration: Supports CUDA, MPS (Apple Silicon), and CPU

Dataset

This project uses the UTKFace dataset, which contains over 20,000 face images with annotations for age, gender, and ethnicity. The images are labeled with ages ranging from 0 to 116 years old.

Dataset Structure: Images are named with the format [age]_[gender]_[ethnicity]_[timestamp].jpg

Age Group Mapping:

• Class 0: Under 16 years old
• Class 1: 16-25 years old
• Class 2: Over 25 years old

Place the UTKFace dataset in the data/UTKFace/ directory.

Installation

Prerequisites

• Python 3.7+
• PyTorch
• OpenCV
• torchvision
• matplotlib
• PIL (Pillow)

Setup

1. Clone the repository:

git clone https://github.com/Kasper-Bankler/AgeClassification.git
cd AgeClassification

2. Install dependencies:

pip install torch torchvision opencv-python matplotlib pillow

3. Download the UTKFace dataset and place it in the data/UTKFace/ directory.

Usage

Training Models

Simple CNN Model

Train the custom CNN model:

python simple_train.py

Configuration:

• Batch size: 32
• Learning rate: 0.0001
• Epochs: 10
• Class weights: [2.0, 2.0, 0.5] (weighted to focus on minors)

The trained model will be saved as simple_model.pth with training metrics visualized in simple_training_metric.png.

Transfer Learning Model

Train the MobileNetV2-based transfer learning model:

python transfer_train.py

Configuration:

• Architecture: MobileNetV2 (pretrained on ImageNet)
• Fine-tuning: Last 4 blocks unfrozen
• Batch size: 32
• Learning rate: 0.0001
• Epochs: 10

The trained model will be saved as transfer_model.pth with training metrics in training_metric.png.

Real-time Inference

Simple Model Inference

Run age classification using the simple CNN model with your webcam:

python simple_camera_inference.py

Make sure your trained model file is in the trained_models/ folder.

Transfer Model Inference

Run age classification using the transfer learning model:

python transfer_camera_inference.py

Controls:

• Press q to quit the inference window

Permissions: On macOS, ensure camera permissions are granted to your terminal/IDE in System Settings → Privacy & Security → Camera.

Model Architectures

Simple CNN

A lightweight custom convolutional neural network:

• Input: 224x224 RGB images
• Feature Extractor: 2 convolutional blocks (3→16→16 channels)
• Pooling: MaxPool2d after each conv block
• Classifier: Fully connected layer with dropout (0.3)
• Output: 3 classes

Transfer Learning (MobileNetV2)

Fine-tuned MobileNetV2 architecture:

• Base Model: MobileNetV2 pretrained on ImageNet
• Fine-tuning: Last 4 blocks trainable
• Custom Classifier:
  • Dropout (0.3)
  • Linear (1280 → 512)
  • ReLU
  • Dropout (0.3)
  • Linear (512 → 3)
• Output: 3 classes

Project Structure

AgeClassification/
├── data/
│   └── UTKFace/              # Dataset directory
├── trained_models/           # Saved model weights
├── plots/                    # Training visualizations
├── dataset.py               # Dataset loader and transforms
├── simple_train.py          # Simple CNN training script
├── transfer_model.py        # Transfer learning model definition
├── transfer_train.py        # Transfer learning training script
├── simple_camera_inference.py   # Real-time inference (simple model)
├── transfer_camera_inference.py # Real-time inference (transfer model)
└── README.md

Training Metrics

Both models track the following metrics during training:

• Training Loss: Cross-entropy loss on training set
• Validation Loss: Cross-entropy loss on validation set
• Illegal Sales Rate: Percentage of under-16 individuals classified as 25+
• Customer Annoyance: Percentage of 25+ individuals flagged as under 25

Class weights are applied to prioritize correct classification of minors to minimize illegal sales risk.

Requirements

• torch - PyTorch deep learning framework
• torchvision - Computer vision utilities and pretrained models
• opencv-python - Real-time computer vision and webcam access
• matplotlib - Training metric visualization
• Pillow - Image processing

Device Support

The code automatically detects and uses the best available hardware:

• CUDA: NVIDIA GPUs
• MPS: Apple Silicon (M1/M2/M3 chips)
• CPU: Fallback for systems without GPU acceleration