Classification of landmine using computer vision

This repository contains code and resources for detecting and classifying land mines. The primary goal is to analyze and classify data derived from ground-penetrating radar and other sensors. The codebase supports machine learning techniques to ensure accurate and efficient classification. This project utilizes a Vision Transformer (ViT) model for classifying land mines using image data. The ViT model, specifically the vit_b_16 variant from the PyTorch torchvision library, is employed with pretrained weights as a feature extractor. The method involves:

1. Preprocessing: Sensor data is preprocessed and resized to match the ViT input dimensions (e.g., 224x224 pixels for images).
2. Feature Extraction: The pretrained ViT processes the input data to extract high-dimensional features using its transformer-based architecture, which excels at capturing long-range dependencies and spatial patterns.
3. Classifier Training: A custom classification head is trained on top of the extracted features to classify the presence of land mines in a binary manner.
4. Evaluation: The model is evaluated using standard performance metrics such as accuracy, precision, and recall, ensuring its robustness and reliability for land mine detection tasks.

Data: Free zone, 0cm depth mine, 1cm depth mine, 5cm depth mine, 10 cm depth mine

Results

• Phase 1 of classification: Free zone vs. Land mine (all mine classes)

Train accuracy: 93%

Test accuracy: 84%

• Phase 2 of classification: Surface mine (0cm & 1cm depth) vs. Deep mine (5cm & 10cm depth)

Train accuracy: 86%

Test accuracy: 75%

Dataset

The data used in this project is sourced from the following publication:

• Title: Land Mine Detection Dataset
• Source: ScienceDirect

The dataset contains sensor readings and annotations that are used for training and validating machine learning models.

Accessing the Dataset

To access the dataset:

1. Visit the dataset link.
2. Follow the instructions on the website to download the data.

Classification model:

The pre-trained ViT classifier was used. Through transfer-learning, all the layers were frozen. Then the head layer was modified for 2 classes and a drop-out with 20% probability.

Project Structure

• Land_mine_main.ipynb: Main notebook containing the code for data preprocessing, model training, evaluation, and visualization.
• data/: Directory where raw and processed data files should be placed (not provided in this repo).
• Land_mine_data_management_functions.py/: Python functions for data management for this project. Data automatically is copied into train and test folders.
• going_modular/data_setup.py/: Python functions to pre-process the images.
• going_modular/engin.py/: Python functions to run the classification model (train and test).

Usage

1. Clone the repository:

   git clone https://github.com/yourusername/land-mine-detection.git
   cd land-mine-detection

2. Place the downloaded dataset in the data/ folder.

3. Open Land_mine_main.ipynb in Jupyter Notebook or any compatible environment.

4. Run the notebook cells step by step to preprocess the data, train the model, and evaluate its performance.

Results

• Model accuracy, precision, and recall metrics are reported in the notebook.
• Visualization of classification results is provided for better interpretability.

Contributing

Contributions are welcome! Please fork the repository and create a pull request with your changes.

License

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

Acknowledgments

• Thanks to the authors of the dataset publication for providing the data.
• Inspired by advancements in ground-penetrating radar technology and machine learning.