🦴 Dinosaur Fossil Classification with ResNet50

Overview

This project explores image classification using ResNet50 through two different approaches:

1. Training from scratch (custom implementation in PyTorch)
2. Transfer learning (pretrained ResNet50 fine-tuned on custom data)

The goal is to understand both the theoretical foundations of deep convolutional networks and the practical advantages of pretrained models.

---

Dataset

• ~1400 images of dinosaur fossils

• Manually collected and organized into class folders

• Categories include examples such as:

  • Sauropod
  • Theropod
  • Ornithiscia
  • Marine(Ammonites, Plesiosaurs, Mosasaurs, Trilobites)
  • Unknown (Eggs, Footprints, Leaves, etc.)

Note: The dataset is relatively small, which significantly impacts model performance when training from scratch.

---

Models

1. ResNet50 (From Scratch)

• Implemented manually in PyTorch
• Includes residual blocks and skip connections
• Trained entirely on the custom dataset

Result:

• Accuracy: 61%

Challenges:

• Limited dataset size
• Slower convergence
• Overfitting and generalization issues

---

2. ResNet50 (Transfer Learning)

• Pretrained on ImageNet
• Fine-tuned on fossil dataset
• Leveraged learned low-level visual features

Result:

• Accuracy: 85%

Advantages:

• Faster convergence
• Better generalization
• More robust feature extraction

---

Model Comparison

Model	Accuracy
From Scratch	61%
Transfer Learning	85%

Key Insight

Transfer learning significantly improves performance when working with limited data by reusing features learned from large-scale datasets.

---

Deployment

The project is deployed as an interactive web application:

• Frontend: Gradio
• Backend: FastAPI
• Infrastructure: AWS ECS Fargate

Features

• Upload an image of a fossil
• Run inference on both models simultaneously
• Compare predictions side-by-side

---

Project Structure

.
├── app/
│   ├── main.py
│   └── example.json
│   └── fun_facts.json
│   └── resnet_finetuned.pth
│   └── resnet_scratch.pth
│   └── torch_utils_finetuned.py
│   └── torch_utils_scratch.py
├── front-end/
│   ├── Dockerfile
│   ├── requirement.txt
│   └── T-Rex-background.jpg
│
├── scratch.ipynb
├── finetuned.ipynb
├── Dockerfile
├── README.md
└── requirements.txt

---

How to Run

1. Install dependencies

pip install -r requirements.txt

2. Run backend

uvicorn app.main:app --reload

3. Launch Gradio UI

python app/gradio_ui.py

---

Future Improvements

• Increase dataset size
• Apply data augmentation techniques
• Experiment with other architectures (EfficientNet, ViT)

---

Acknowledgments

• PyTorch
• ImageNet pretrained models
• Open-source fossil image sources

---

Author

Somprat Suratannon

License

MIT License