This project implements a deep learning pipeline for road skeletonization from aerial imagery, using a customized Efficient U-Net model. The system includes data preparation, training, evaluation, and inference scripts.

• Road Skeletonization with Efficient U-Net
  • Table of Contents
  • Project Overview
  • Installation
  • Usage
    • Train Model
    • Test Model
  • Model Architecture
  • File Descriptions
  • License
  • Notes

This project aims to generate 1-pixel-wide road skeleton maps from input aerial images. It uses a 5-down Efficient U-Net structure with Leaky ReLU activations and Batch Normalization layers.

You can install the required dependencies using the provided installation script:

bash setup.sh

To train the model:

python train.py

Training outputs model checkpoints to the saved_models/ directory.

To test the model on unseen data:

python test.py

The predictions will be saved under a results/ directory.

The model is based on a simplified U-Net design with the following structure:

• Double Convolution blocks (Conv2D -> BatchNorm -> LeakyReLU -> Conv2D -> BatchNorm -> LeakyReLU)
• 5 levels of downsampling and upsampling
• Skip connections between matching resolution levels

Implemented in model.py.

This project is released under the MIT License.

• Make sure your dataset is structured correctly (input images and corresponding masks).
• CUDA-enabled GPU is recommended for faster training.
• The project assumes input and output images are 256x256 pixels.