FRESCO: Factorized Representation for Enhanced Super-resolution using latent Component-adversarial Optimization

This repository corresponds to the paper:

Unregistered Spectral Image Fusion: Unmixing, Adversarial Learning, and Recoverability Jiahui Song, Sagar Shrestha, Xiao Fu

[ arXiv ]

Overview

FRESCO addresses the unregistered hyperspectral and multispectral image fusion problem using a two-stage framework:

MSR stage

Coupled spectral unmixing to recover the M-SRI.

HSR stage

Latent-space adversarial learning to recover the H-SRI by matching abundance patch distributions.

The method does not require paired training data or spatial co-registration.

Reproduce the Experiment

1. Installation

The algorithm was implemented with Python 3.12.8. Please install the dependencies as follows:

conda create -n fresco python=3.12.8
conda activate fresco
pip install -r requirements.txt

2. Data Preparation

(1) Download the dataset from the link: [DatasetLink].

(2) After downloading, unzip the dataset and place it under the data/ folder.

3. Run the Code

Activate the environment and run the corresponding experiment script. For example, to reproduce the experiment on the Pavia University dataset with Gaussian-kernel downsampling, run:

conda activate fresco
bash scripts/pavia_gaussian.sh

Semi-real Experiments

MSR Results

HSR Results

Real Applications

Citation

If you use this work, please cite:

@article{song2026unregistered,
  title={Unregistered spectral image fusion: Unmixing, adversarial learning, and recoverability},
  author={Song, Jiahui and Shrestha, Sagar and Fu, Xiao},
  journal={arXiv preprint arXiv:2603.21510},
  year={2026}
}