README.md

---
title: Age Detection ResNet50
emoji: 👨‍👩‍👧‍👦
colorFrom: blue
colorTo: green
sdk: gradio
sdk_version: 4.44.0
app_file: app.py
pinned: false
license: mit
models:
- Sharris/age-detection-resnet50-model
datasets:
- UTKFace
---

# Age Detection with ResNet50

This Hugging Face Space demonstrates age prediction from facial images using a ResNet50-based regression model trained on the UTKFace dataset.

## 🎯 **Key Features**

- **Advanced Architecture**: ResNet50 backbone with 256×256 input resolution
- **Bias Correction**: Inverse frequency sample weighting to address dataset imbalances
- **Robust Training**: Huber loss for outlier resilience and label noise tolerance
- **High Accuracy**: 19.96 years validation MAE with comprehensive bias correction

## 🔧 **Model Details**

- **Architecture**: ResNet50 pre-trained on ImageNet
- **Input Size**: 256×256×3 RGB images
- **Loss Function**: Huber loss (δ=1.0) for robust regression
- **Sample Weighting**: Inverse frequency weighting (0.225x to 34.259x by age)
- **Training Data**: UTKFace dataset with age range 0-116 years

## 💡 **Usage Tips**

1. **Crop to face**: For best results, crop the image to show mainly the face
2. **Good lighting**: Ensure the face is well-lit and clearly visible
3. **Front-facing**: Works best with front-facing portraits
4. **Age range**: Trained on ages 0-116, most accurate for common age ranges

## 📊 **Performance**

- **Validation MAE**: 19.96 years
- **Bias Correction**: Addresses systematic young-age bias through sample weighting
- **Age Distribution**: Balanced training across all age groups using inverse frequency weighting

## 🔬 **Technical Implementation**

- Built with TensorFlow/Keras 3
- ResNet50 preprocessing pipeline
- Gradio interface for easy web deployment
- Automatic model download from Hugging Face Hub

Upload a face image and get an instant age prediction!