khlmtf / segmantaion_roofs_satellite Goto Github PK

// README.txt

# Roof Segmentation Project

This project involves developing a U-Net model for segmenting roofs in satellite images. The goal is to accurately identify and segment roofs from a set of aerial images.

## Project Structure

- **data/**: Contains the input images and labels.
- **models/**: Directory where the trained model is saved.
- **predictions/**: Directory where the predicted masks and visualizations are saved.
- **scripts/**: Contains the preprocessing, training, and evaluation scripts.
  - `preprocess_data.py`: Script for preprocessing the input data.
  - `train_model.py`: Script for training the U-Net model.
  - `evaluate_model.py`: Script for evaluating the trained model and visualizing the results.

## Requirements

The project requires the following packages:

- numpy
- opencv-python
- tensorflow
- matplotlib
- scikit-learn

You can install the necessary packages using the following command:

```bash
pip install -r requirements.txt

Usage
Preprocess the Data

Before training the model, preprocess the data by running:
python scripts/preprocess_data.py

- Train the Model, to train the U-Net model, run:
python scripts/train_model.py

This will train the model and save it in the models directory.
Evaluate the Model, to evaluate the trained model and visualize the predictions, run:
python scripts/evaluate_model.py

This will save the predicted masks and the visualization plot in the predictions directory.
Results

The evaluation script provides the following results:
    Test Loss: 0.303048
    Test Accuracy: 0.855072

The predicted masks have been saved in the predictions directory, along with a visualization plot (predictions/predictions.png).
Visualizations

The visualizations show the input images, ground truth masks, and the predicted masks side by side.
Acknowledgments

This project uses the U-Net architecture for image segmentation, which is widely used for biomedical image segmentation and has proven to be effective for other types of image segmentation tasks.