PT-Fusion

PT-Fusion: A Multi-Modal Attention Network for Defect Analysis in Pulse Thermography

You can find the PDF of the paper here. If you use this code in an academic context please cite this publication:

@misc{pt_fusion,
      title={Multi-Modal Attention Networks for Enhanced Segmentation and Depth Estimation of Subsurface Defects in Pulse Thermography}, 
      author={Mohammed Salah and Naoufel Werghi and Davor Svetinovic and Yusra Abdulrahman},
      year={2025},
      eprint={2501.09994},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2501.09994}, 
}

Code Structure and IRT-PVC Dataset Outline:

Supported platforms

Tested on the following platforms:

• Ubuntu 18.04 and 20.04 LTS

Prerequisites

You need the model checkpoints and dataset to be in your working directory:

Checkpoints Testing Dataset

Running the code

Step 1: Create a conda environment

Create pt_fusion conda environment:

git clone https://github.com/mohammedsalah98/pt_fusion.git
cd pt_fusion
conda env create -f environment.yml
conda activate pt_fusion

Step 2: Running the code:

Multi-Class Segmentation:

cd pt_fusion
python test_multi.py --checkpoint /path/to/checkpoint --data_folder /path/to/dataset

Segmentation & Depth Estimation:

python test_depth.py --checkpoint /path/to/checkpoint --data_folder /path/to/dataset

Parameters:

• --checkpoint: Path to downloaded checkpoint
• --data_folder: Path to dataset

Training:

Training Dataset

python train_segmentation.py --data_folder /path/to/dataset
python train_depth.py --data_folder /path/to/dataset

For example, if you place the checkpoint in the checkpoints folder and the dataset in pt_fusion directory, the command should look like this:

python test_depth.py --checkpoint checkpoints/attention_fusionUnet_depth.pth --data_folder dataset/

Disclaimer

We will soon release a code for benchmarks against state-of-the-art models.