Water-KAN: Efficient Underwater Image Enhancement via Kolmogorov–Arnold Networks
IEEE Signal Processing Letters · 2026 · 33 · 2810-2814
DOI: 10.1109/lsp.2026.3708416 ↗Kurz gefasst
Fachartikel in der Zeitschrift IEEE Signal Processing Letters, 2026. Verfasst von Peiyuan Jin und 3 weiteren Autor:innen. Die Arbeit nennt 27 Literaturangaben. Gefördert von: National Natural Science Foundation of China, Zhejiang Natural Science Foundation of China.
Abstract (Originalsprache)
Underwater image enhancement (UIE) is challenging due to wavelength-dependent attenuation causing non-linear color shifts and contrast loss. Traditional CNNs often struggle here, as their static activations lack the flexibility to model exponential light decay. We introduce Water-KAN, a hybrid architecture leveraging Kolmogorov–Arnold Networks for precise restoration. Its core, the KAN Bottleneck, cascades RBF-KAN Blocks utilizing learnable Gaussian activations. This design approximates intricate non-linear mappings more effectively than standard convolutions. Furthermore, we develop a Depth-SFT Decoder that leverages depth priors to enforce geometric consistency across scales. Extensive tests on UIEB and LSUI benchmarks demonstrate Water-KAN’s superiority. Notably, it improves PSNR by 1.941 dB over comparable CNN baselines, achieving a top-tier SSIM of 0.912.
Förderung
- National Natural Science Foundation of China
- Zhejiang Natural Science Foundation of China
Zitieren
APA
Jin, P., Xu, H., Chen, Y., & Jiang, G. (2026). Water-KAN: Efficient Underwater Image Enhancement via Kolmogorov–Arnold Networks. IEEE Signal Processing Letters, 33, 2810-2814. https://doi.org/10.1109/lsp.2026.3708416
Vancouver
Jin P, Xu H, Chen Y, Jiang G. Water-KAN: Efficient Underwater Image Enhancement via Kolmogorov–Arnold Networks. IEEE Signal Processing Letters. 2026;33:2810-2814. doi:10.1109/lsp.2026.3708416
BibTeX
@article{Jin2026,
author = {Jin, Peiyuan and Xu, Haiyong and Chen, Yeyao and Jiang, Gangyi},
title = {Water-KAN: Efficient Underwater Image Enhancement via Kolmogorov–Arnold Networks},
journal = {IEEE Signal Processing Letters},
year = {2026},
volume = {33},
pages = {2810--2814},
doi = {10.1109/lsp.2026.3708416}
}
EUVIMED