Repository for the dataset and code associated with the paper "Interpretability of Deep High-Frequency Residuals: a Case Study on SAR Splicing Localization".
The repository is currently under development, so feel free to open an issue if you encounter any problems.
This project provides tools and data to analyze the interpretability of high-frequency residuals in deep learning models applied to manipulation localization in SAR (Synthetic Aperture Radar) images. It includes scripts for patch extraction, network models, data management utilities, and analysis notebooks.
extract_dhfrs.py: Main script for high-frequency residual extraction.isplutils/: Python modules for data management, neural networks, and patch extraction.data/poc/: Examples of manipulated SAR images and reference masks.weights/: Pre-trained model weights.notebooks/: Jupyter notebooks for analysis and proof of concept.
To install dependencies, use the environment.yaml file:
conda env create -f environment.yaml
conda activate dhfr_interpretability- Download the dataset from this link (coming soon!) and place it in the
data/directory. - Extract patches and residuals with
extract_dhfrs.py. - Analyze results using the notebooks in
notebooks/. - Check the modules in
isplutils/for customization or extensions.
A proof of concept is available in the notebooks/ directory, demonstrating the interpretability analysis on a sample SAR image showed in the paper.