Domain Decomposition-Inspired CNN for Image Segmentation: Leveraging Multiple GPUs for High-Resolution Images

Welcome to the repository for the thesis titled “A Domain Decomposition-based CNN Architecture for High-Resolution Image Segmentation.” This repository contains the code scripts used to generate the results presented in the thesis. Below is a brief overview of the repository structure:

• data: This folder is dedicated to storing synthetic and realistic datasets.
• dataset_tools: Contains several useful scripts for initializing datasets, dataloaders, data transformations, processing data, and functions to generate synthetic datasets.
• models_2d: This directory holds Torch scripts initializing the model proposed in the thesis, along with a baseline U-Net model and various model components such as the encoder, decoder, and communication network.
• evaluate.py: Contains functions used to evaluate model metrics by generating a confusion matrix.
• train_deepglobe_resnet_unet.py: Holds the function used to train a ResNet-UNet model for image segmentation on the DeepGlobe landtype dataset.

1. Data Preparation: Place your datasets in the data folder or use the provided scripts in dataset_tools to generate synthetic datasets.
2. Model Initialization: Access the scripts in the models_2d folder to initialize the proposed model or baseline models.
3. Training: Utilize train_deepglobe_resnet_unet.py to train your segmentation model.
4. Evaluation: After training, use evaluate.py to evaluate model performance and generate confusion matrices.