This repository contains the codebase and supporting materials for the master's thesis "High-resolution Registration of Irradiated Concrete Scans" by Erik Philippe Leclercq. It contains a complete image processing pipeline for the registration of high-resolution video scans of rotating concrete samples that were extracted from nuclear power plants.

There is a jupyter notebook available at src/demo.ipynb that goes through the video processing pipeline step by step. In order to be able to use it, you will need to create a Python venv, download the required packages, and download the dataset (or parts of it).

Description of the repository for future developers is given in contributions.md.

The designed pipeline enables:

• Remote analysis of radioactive concrete samples.
• Automatic video registrationg using state-of-the-art algorithms.
• Generation of composite images for visual inspection and automated crack detection.

The pipeline is fully configurable and modular. Modules in src/ contain configuration files and a validation schema that ensure that the processed config file is valid. The config files are well-documented with explanations of what each parameter controls.

• 4k video scans of cylindrical concrete samples (~390 degree rotation) that contain two illumination conditions.
• Two illumination conditions: low-angle (≤15°) and front-lighting (≥82°).

1. Video splitting: into light-specific segments.
2. Rotation correction via optical flow: and other video processing steps.
3. Video registration with the specified method from one of the following choices: ['mudic', 'orb', 'lightglue'].
4. Composite image creation with the use of ['max', 'min', 'mean'] image fusion.
5. Evaluation of composite images with the normalised grey-level variance and the Brenner's evaluation metrics.

Structure of the pipeline is described in src/README.md

• Python 3.11
• GPU with cuda support
• (optional) 50GB of disc space for the dataset
• FFmpeg, which can be downloaded here.. You will also need to download the libx264-dev library from here.
• git lfs installed, instruction on how to install git lfs can be found here.
• jupyter, can be install from here.

git clone https://github.com/research-centre-rez/leclercq-thesis.git
cd leclercq-thesis

pip install .

# Create the jupyter notebook demo
jupytext src/demo.py --to ipynb

The dataset can be downloaded with the following bash script:

git clone https://huggingface.co/datasets/research-centre-rez/concrete-samples
cd concrete-samples
git lfs pull

The program is a cli-based tool that you can use globally, the list of available commands is:

1. cst-pre-processing: Removes periods of darkness from the video.
2. cst-video-processing: Prepares the input video for registration.
3. cst-video-register: Registers a video sequence into an .npy file.
4. cst-image-fusion: Image fusion of the registered samples.
5. cst-image-evaluation: Evaluates the quality of the registration.
6. cst-create-before-after-pairs: Pairs up before and after scans of the sample radiation exposure.
7. cst-crack-analysis: Identifies cracks in the input samples.

Each command is well documented, simply accessible with cst-command --help

For each individual video, the pipeline outputs the following:

• The registered stack as a .npy file, which essentially contains the whole video decompressed into its individual frames.
• One or more of the specified composite images.
• Logs of all the applied transformations.
• Evaluation of the composite images with the specified metrics.
• In case of registration failure a src/failures.log is created that logs what exactly went wrong.

If you use this work, please cite:

@mastersthesis{leclercq_thesis_2025,
    author = {Erik Philippe Leclercq},
    title = {High-Resolution Registration of Irradiated Concrete Scans},
    school = {Charles University},
    year = {2025},
}