README.md

metadata

title: Astronomy Image Classification
emoji: 🌌
colorFrom: blue
colorTo: purple
sdk: streamlit
sdk_version: 1.28.0
app_file: app.py
pinned: false
license: mit

🌌 Astronomy Image Classification - Ensemble Model

A deep learning ensemble system for classifying astronomy images into 6 categories using ResNet50 and DenseNet121 models with soft voting.

Model Performance

• ResNet50 Accuracy: 64.86%
• DenseNet121 Accuracy: 63.96%
• Ensemble Expected Accuracy: 70-75%
• Target Accuracy: >95%
• Architecture: ResNet50 + DenseNet121 Ensemble
• Framework: PyTorch
• Input Size: 224x224 pixels

Ensemble Method

This system uses soft voting to combine predictions from both models:

1. Each model makes independent predictions
2. Probabilities are averaged across models
3. Final prediction is the class with highest average probability
4. Provides higher accuracy than individual models

📊 Classes

1. 🌟 Constellation - Star patterns forming recognizable shapes (Orion, Big Dipper)
2. ** Cosmos** - General space scenes and cosmic phenomena
3. ** Galaxies** - Spiral, elliptical, and irregular galaxies (Andromeda, Milky Way)
4. 💫 Nebula - Gas clouds and stellar nurseries (Orion Nebula, Eagle Nebula)
5. 🪐 Planets - Solar system planets and planetary features (Jupiter, Saturn, Mars)
6. ⭐ Stars - Individual stars and stellar objects

🚀 Usage

1. Upload an astronomy image (JPG, PNG, JPEG)
2. View individual model predictions
3. See ensemble prediction with confidence scores
4. Explore all class probabilities

🔧 Technical Details

• Models: ResNet50 (95MB) + DenseNet121 (30MB)
• Preprocessing: Resize to 224x224, ImageNet normalization
• Augmentation: Albumentations library
• Optimization: AdamW with cosine scheduling
• Loss Function: CrossEntropy with class weights
• Ensemble: Soft voting (average probabilities)

📈 Individual Model Results

Model	Accuracy	Precision	Recall	F1-Score
ResNet50	64.86%	0.6594	0.6486	0.6452
DenseNet121	63.96%	0.6461	0.6396	0.6172
Ensemble	~70%	Higher	Higher	Higher

🎨 Sample Images

Upload images of:

• Constellations: Star patterns, asterisms
• Galaxies: Spiral, elliptical, irregular galaxies
• Nebulae: Emission, reflection, dark nebulae
• Planets: Solar system planets, planetary features
• Stars: Individual stars, stellar phenomena
• Cosmos: Deep space, cosmic phenomena

🚀 Deployment Features

• ✅ Interactive Web Interface - Easy image upload
• ✅ Real-time Predictions - Instant classification
• ✅ Ensemble Results - Both individual and combined predictions
• ✅ Confidence Scores - Visual confidence indicators
• ✅ All Class Probabilities - Complete probability breakdown
• ✅ Mobile Friendly - Responsive design
• ✅ Error Handling - Robust error management

🔮 Future Improvements

• Test Time Augmentation (TTA) - Multiple augmented predictions
• More Models - Add EfficientNet, Vision Transformer
• Advanced Ensemble - Weighted voting based on performance
• Progressive Training - Multi-stage training approach
• Data Augmentation - More aggressive augmentation
• Transfer Learning - Pre-training on larger datasets

##️ Local Testing

# Install dependencies
pip install -r requirements.txt

# Run locally
streamlit run app.py

📁 Model Files

• best_resnet50.pth - ResNet50 model weights (95MB)
• best_densenet121.pth - DenseNet121 model weights (30MB)
• model.py - Model architecture definition
• inference.py - Inference pipeline with ensemble
• app.py - Streamlit web application

---

• Built with ❤️ for astronomy enthusiasts and data scientists*
  🎯 Target: >95% accuracy through ensemble methods and advanced techniques