IEEE Transactions on Biomedical Engineering 2026
ZS-KAN
ZS-KAN: Zero-Shot Image Denoising With Lightweight Kolmogorov–Arnold Networks
Current or former lab member. Select a name to see all of their projects.
Abstract
Learning-based denoisers usually depend on deep architectures and large datasets, which limits them where data are scarce or computing is constrained. ZS-KAN is a self-supervised, zero-shot alternative: it denoises a single noisy image without any training set, using a hybrid of convolutional layers and Kolmogorov–Arnold networks (KANs).
On synthetic and real noisy images, ZS-KAN performs comparably to or better than state-of-the-art zero-shot denoisers while using only 1% to 25% of their parameters, and it typically preserves more fine detail. The small footprint makes it a practical candidate for medical imaging workflows and edge devices.
What’s in the repository
- A unified command-line tool,
zskan, for single-image denoising and dataset-level evaluation - Network source code, scripts and tests
- Instructions for data preparation
Citation
@article{wang2026zskan,
title = {ZS-KAN: Zero-Shot Image Denoising With Lightweight Kolmogorov–Arnold Networks},
author = {Wang, Jianxu and Wang, Ge},
journal = {IEEE Transactions on Biomedical Engineering},
volume = {73},
number = {7},
pages = {2602--2611},
year = {2026},
doi = {10.1109/TBME.2025.3640551}
}