Image segmentation with UNet model

Data, predictions and saved model can be found here: https://drive.google.com/drive/folders/1eFxwcWIMcmCI9F7FzI-nLLcA9-BBp-Cq

Introduction

This project focuses on semantic segmentation of images using the U-Net architecture.

Prerequisites

• Python: 3.10
• Libraries: TensorFlow 2.x, NumPy, Matplotlib, Pillow

Model Architecture

The U-Net architecture is a type of convolutional neural network that is widely used for semantic segmentation tasks. Its architecture is symmetrical and consists of an encoder and a decoder part, which are connected by a bottleneck.

Details of the architecture can be found here.

Project Structure

• dataset.py: Implements DatasetGenerator class which generates batches of images from the dataset.
• labeldata.py: Defines attributes and colors for cityscape image segmentation labels and maps label IDs to their respective colors.
• model.py: Defines the UNet model architecture.
• predictions.py: Makes predictions on the validation dataset and saves them to predictions/ directory.
• train.py: Trains the model on training data and saves model.
• results.ipynb: Displays a small sample of predictions.

Results

Predictions are made on the validation dataset and the results are saved in the drive link provided above. A small sample of these predictions can be seen in the results.ipynb file.