A deep learning project classifying human faces into age groups using transfer learning with CNN architectures, achieving 72.38% accuracy with Grad-CAM visualizations for model interpretability.
The Challenge
Age estimation from facial images is challenging due to high variability in pose, expression, illumination, and ethnicity. Manual age verification is slow and inconsistent, while automated systems often lack interpretability.
The Solution
Built and compared three CNN architectures (MobileNetV2, ResNet50, EfficientNet-B0) using transfer learning with ImageNet-pretrained weights. ResNet50 achieved the best performance at 72.38% accuracy on 4-class age classification, with Grad-CAM visualizations explaining model decisions.
Key Features
- Transfer learning with ImageNet-pretrained MobileNetV2, ResNet50, and EfficientNet-B0
- 4-class age classification: Young (0-25), Adult (26-50), Middle-aged (51-75), Senior (76+)
- Grad-CAM visualizations highlighting facial regions used for predictions
- Custom classifier head with dropout regularization
- Google Colab integration for accessible GPU training
Tech Stack
Deep learning framework with torchvision for pre-trained models and image transforms
Jupyter Notebook environment for iterative development with 20,000+ face images
Model evaluation metrics including confusion matrices and accuracy scoring
Image processing for Grad-CAM heatmap generation and visualization