🐾 Animal Image Classifier

Multi-class animal image classification system achieving 98.24% test accuracy using EfficientNet-B3 with transfer learning.

🚀 Try the Live Demo

🎯 Overview

The system classifies images across 10 animal classes with state-of-the-art accuracy.

Key Features

• 98.24% test accuracy on 3,928 test images
• Grad-CAM visualization for model interpretability
• Confidence-based automation tiers for deployment
• Side-by-side model comparison (EfficientNet vs ResNet50)
• Interactive web demo built with Gradio

📊 Results

Model	Parameters	Test Accuracy	Key Insight
EfficientNet-B3 ✅	10.7M	98.24%	Winner: Best accuracy with 58% fewer params
ResNet50	25.6M	97.0%	Strong baseline, less efficient
ViT-Base	85.8M	96.51%	Underperforms on medium datasets
Baseline CNN	0.4M	50.0%	Training from scratch fails

Key Finding: Transfer learning provides 48% accuracy improvement over training from scratch.

🛠️ Technologies

• Framework: PyTorch
• Architecture: EfficientNet-B3 (pre-trained on ImageNet)
• Interpretability: Grad-CAM
• Interface: Gradio
• Deployment: Hugging Face Spaces

📁 Project Structure

animal-image-classifier/
├── app.py                          # Gradio web application
├── requirements.txt                # Python dependencies
├── examples/                       # Example images for demo
└── README.md                       # This file

🚀 Quick Start

Installation

# Clone the repository
git clone https://github.com/Omar-Camara/animal-image-classifier.git
cd animal-image-classifier

# Create virtual environment
python -m venv venv
source venv/bin/activate  # On Windows: venv\Scripts\activate

# Install dependencies
pip install -r requirements.txt

Download Model Files

Model files are too large for GitHub. Download them from the Hugging Face Space:

1. Download efficientnet_b3_best.pth (~43 MB)
2. Download resnet50_finetuned_best.pth (~103 MB)
3. Place both files in the project root directory

Or download directly:

# Download EfficientNet model
curl -L -o efficientnet_b3_best.pth "https://huggingface.co/spaces/Username273183/animal-classifier/resolve/main/efficientnet_b3_best.pth"

# Download ResNet model
curl -L -o resnet50_finetuned_best.pth "https://huggingface.co/spaces/Username273183/animal-classifier/resolve/main/resnet50_finetuned_best.pth"

Run the Demo

python app.py

Then open http://localhost:7860 in your browser.

📖 How It Works

1. Transfer Learning

• Pre-trained EfficientNet-B3 on ImageNet (1.2M images, 1000 classes)
• Fine-tuned on animal dataset (26,179 images, 10 classes)
• Achieves 48% improvement over training from scratch

2. Grad-CAM Visualization

• Visualizes which parts of the image the model focuses on
• Confirms model learns anatomically relevant features
• No background bias or spurious correlations

3. Confidence-Based Automation

• ≥90% confidence: Auto-accept (92% of predictions, ~100% accuracy)
• 70-90% confidence: Flag for review (5% of predictions)
• <70% confidence: Require verification (3% of predictions)

🔬 Methodology

Dataset

• Total Images: 26,179
• Classes: butterfly, cat, chicken, cow, dog, elephant, horse, sheep, spider, squirrel
• Split: 70% train / 15% validation / 15% test
• Class Imbalance: 3.36:1 ratio (handled naturally by transfer learning)

Training Configuration

• Epochs: 5 (transfer learning converges quickly)
• Optimizer: Adam (lr=0.0001)
• Batch Size: 64
• Data Augmentation: Horizontal flips, rotation (±15°), color jitter
• Hardware: Tesla T4 GPU
• Training Time: ~70 minutes

Evaluation Metrics

• Test Accuracy: 98.24% (3,859/3,928 correct)
• Macro F1-Score: 98.01%
• All classes: >96% F1-score
• Zero errors above 90% confidence threshold

📈 Key Research Findings

1. Transfer learning is essential

   • 48% accuracy gain vs training from scratch (50% → 98%)
   • Converges in 2 epochs vs 20+ for baseline

2. Architecture efficiency matters

   • EfficientNet (10.7M params) beats ViT (85.8M params)
   • Systematic scaling > brute force parameter increase

3. Class imbalance handling not always needed

   • Moderate imbalance (3.36:1) handled naturally
   • Weighted/focal loss decreased performance

4. Data quality limits performance

   • ~25% of errors traced to dataset mislabeling
   • Model confidence serves as quality detector

5. Uncertainty quantification enables deployment

   • Clear confidence separation (95% vs 69%)
   • Enables 92% automation with >99% accuracy

📄 Citation

If you use this work, please cite:

@misc{camara2024animal,
  title={Multi-Class Animal Image Classification Using Transfer Learning: 
         A Comparative Study of Deep Learning Architectures},
  author={Camara, Omar},
  year={2025},
  institution={Syracuse University},
}

👨‍💻 Author

Omar Camara
Graduate Student, Computer Science
Syracuse University

• 📧 Email: omarcamara000@gmail.com
• 💼 LinkedIn: linkedin.com/in/oc18
• 🐙 GitHub: github.com/Omar-Camara

🎓 Project Context

This project was completed as part of coursework at Syracuse University (Fall 2025). The work demonstrates:

• Systematic experimental methodology
• Transfer learning effectiveness
• Production deployment considerations
• Model interpretability and uncertainty quantification

📝 License

This project is licensed under the MIT License - see the LICENSE file for details.

Acknowledgments

• EfficientNet and ResNet architectures from torchvision
• Grad-CAM implementation inspired by original paper
• Gradio team for the excellent web framework

🔗 Links

• 🚀 Live Demo
• 📊 Hugging Face Space

---

⭐ Star this repo if you find it helpful!