Oxford Pet Classification with SHAP and LIME

This project focuses on image classification using a pre-trained convolutional neural network and interprets the model's predictions using SHAP (SHapley Additive exPlanations) and LIME (Local Interpretable Model-Agnostic Explanations). The project leverages the Oxford-IIIT Pet Dataset for multi-class classification, demonstrating how to train a robust image classifier and explain its predictions using interpretable AI methods.

---

Features

1. Image Classification:

   • Utilizes a pre-trained convolutional neural network (e.g., ResNet, VGG, or another backbone) to classify pet images into multiple categories.

2. Model Training and Validation:

   • Includes training a model on the Oxford-IIIT Pet Dataset with appropriate data augmentation and validation strategies.

3. Explainability:

   • Implements SHAP and LIME to provide detailed insights into the model's predictions:
     • SHAP highlights pixel importance for predictions.
     • LIME explains predictions by locally perturbing image inputs.

4. Visualization:

   • Provides clear visualizations of model predictions and the explanations generated by SHAP and LIME.

---

Dependencies

To run the project, ensure the following libraries are installed:

• Python 3.x
• TensorFlow or PyTorch (depending on the model implementation)
• NumPy
• Matplotlib
• SHAP
• LIME
• scikit-learn
• OpenCV (optional, for image processing)

---

Usage

1. Clone the repository:

   git clone https://github.com/VivianeLe/Image_Classification_SHAP_Lime.git
   cd Image_Classification_SHAP_Lime

2. Open the Jupyter Notebook:

   jupyter notebook OxfordPet-Classification.ipynb

3. Run the cells sequentially to:

   • Load and preprocess the Oxford-IIIT Pet Dataset.
   • Train the classification model.
   • Visualize SHAP and LIME explanations for selected predictions.

---

Results

• Classification Accuracy: The model achieves high accuracy on the validation set, showcasing its ability to classify pet images correctly.
• Explainability Outputs:
  • SHAP produces heatmaps highlighting important image regions influencing the model's decisions.
  • LIME generates perturbed explanations for individual image predictions.

---

References

• Oxford-IIIT Pet Dataset
• SHAP Documentation
• LIME Documentation
• Relevant research papers and blogs on model interpretability.

---

Contributing

Contributions are welcome! Feel free to submit issues or pull requests for improvements or new features.

---

License

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