Deep Learning-Based Learned Skeletonization of Road Networks

This project aims to reconstruct clean, one-pixel-wide skeleton representations from thick and noisy road images. It is part of our coursework for CSE627 - Spring 2024–25, focusing on semantic segmentation using deep learning.

📌 Objective

To develop a U-Net-based model that can extract accurate road skeletons from distorted road network images generated using OpenStreetMap (OSM) data.

🧱 Model Architecture

• Based on U-Net with 5 encoder and 4 decoder levels.
• Enhanced with double convolutional blocks and skip connections.
• Final output: 1x256x256 binary segmentation mask.

🧪 Dataset and Preprocessing

• Input: Thickened 256×256 grayscale road images from OSM.
• Target: One-pixel-wide ground truth skeleton masks.
• Distortions Applied: Gaussian noise, salt & pepper noise, motion blur, and defocus blur.
• Dataset prepared using a custom RoadSkeletonDataset class.

⚙️ Training Details

• Loss Functions: BCEWithLogitsLoss, Dice Loss, and a Combined Loss (weighted sum of both).
• Optimizers: Adam (lr = 1e-4 to 1e-5).
• Batch Size: 8
• Epochs: Up to 250
• Validation Strategy: 70/15/15 train-val-test split with early stopping and learning rate scheduling.

📈 Evaluation Metrics

• Pixel-Level: Test loss, MSE, IoU, Dice coefficient.
• Node-Level: Precision & Recall for valence-1 to valence-4 nodes using bipartite matching.

🔬 Ablation Study

We compared performance using different loss functions and training durations:

• Loss types: BCE, Dice, Combined (BCE + Dice)
• Epochs: 60, 80, 180, and 250
• Observed that Combined Loss at 180+ epochs provided best trade-off between detail preservation and generalization.

🔍 Results

• Best Dice Score: 0.8703
• Best IoU: 0.7894
• Strong node-level recall across valence types, especially valence-2 and valence-4 junctions.
• Visual inspections confirm the continuity and thinness of the predicted skeletons.

📁 Repository Structure

├── dataset_preperation.py        # Dataset loading and distortion augmentation
├── model_train_eval.py           # Training, evaluation, and loss comparison
├── model.py                      # U-Net architecture
├── README.md                     # This file
└── sample_prediction.png         # Example output image (input, pred, ground truth)

🤝 Team Members

• Md Aibak Aljadayah
• Md Nahid Hasan
• Md Nadim Mahmud

🔗 GitHub Repo

https://github.com/aibak365/Road_Skeletonization