Building Classification using ML & Deep Learning

Project Overview

This project aims to classify buildings into five distinct categories (A, B, C, D, and S) using a combination of Machine Learning and Deep Learning techniques. We have utilized traditional ML classifiers (SVM, Logistic Regression, Random Forest, XGBoost) alongside a Neural Network model to achieve robust classification. The project also involves feature extraction, PCA dimensionality reduction, and ensemble techniques to enhance performance.

Table of Contents

• Project Overview
• Dataset
• Approach
• Installation
• Usage
• Results
• Observations and Findings
• Future Work
• Contributions
• License

Dataset

The dataset contains grayscale images of various buildings belonging to five categories (A, B, C, D, S). Each image has a size of (300, 400), with a total of 2466 images for training and testing. The dataset is split into training and test sets with the following distribution:

• A: 289 images
• B: 352 images
• C: 721 images
• D: 904 images
• S: 200 images

Approach

Our approach includes the following key steps:

1. Exploratory Data Analysis (EDA): Initial analysis of the dataset, class distribution, and image properties.
2. Feature Extraction: Extracting features from images using pre-trained efficientNet model.
3. Dimensionality Reduction: Applying PCA to reduce feature dimensions while retaining 95% variance.
4. Model Training: Training classifiers including SVM, Logistic Regression, Random Forest, XGBoost, and a Neural Network.
5. Model Ensembling: Combining the best models using soft voting for improved accuracy.
6. Evaluation: Using a test set for final performance evaluation, including confusion matrices and classification reports.

Installation

To run this project locally, follow the steps below:

1. Clone the repository:

git clone https://github.com/your-username/building-classification.git
cd building-classification

2. Install the dependencies:

Use the provided requirements.txt file to install all necessary packages:

pip install -r requirements.txt

3. Download the dataset:

You can find the dataset here: https://drive.google.com/drive/folders/1RqHqUA83mm-CdqWZxiJ_JrlthtF7EUkX?usp=sharing

4. Run the Jupyter Notebook:

You can view the complete analysis and training process in the provided Jupyter notebook:

jupyter notebook building-classification.ipynb

Usage

1. Training Models:

The models are trained using PCA-reduced features. You can modify hyperparameters in the notebook or directly run the cells to train the models.

2. Making Predictions:

To predict the category of a building image, use the pre-trained models: (as shown at the end of the notebook)

# Load the voting classifier and predict on new images
predicted_label, individual_predictions = predict_image(image_path, pca, classifiers, transforms, device)
print(f"Predicted label: {predicted_label}")

3. Evaluation:

Evaluate the ensemble model using:

print(classification_report(y_test, predictions))
print(confusion_matrix(y_test, predictions))

Results

(more results and visualizations are inside 'results' folder) Our ensemble model achieved X% accuracy on the test dataset, with a detailed classification report provided below. The confusion matrix highlights the model's performance across all categories.

Confusion Matrix:

Classification Report:

Observations and Findings

• EfficientNet was the most suitable model for our small dataset, particularly with data augmentation and advanced loss functions.
• Handling Imbalanced Data: Using Focal Loss and stratified sampling was crucial for dealing with imbalanced classes.
• PCA allowed us to uncover the non-linear structure of the data, leading to significant performance improvements with non-linear classifiers.

Future Work

• Hyperparameter Tuning: Further tuning of the Neural Network and other models could yield even better results.
• Deploying the Model: Building a web application or API to serve the trained models for real-time predictions.

Contributions

Feel free to open issues or pull requests to contribute to the project. All contributions are welcome!

License

This project is licensed under the MIT License.