EUVIMEDEUVIMED entdecken
Crossref

6‐2: Beyond MLPs: Convolutional Color Constancy with Kolmogorov‐Arnold Networks

Liangwei Chen, Ming Ronnier Luo

SID Symposium Digest of Technical Papers · 2026-08 · 57 (S1) · 45-48

DOI: 10.1002/sdtp.70003_12 ↗

Kurz gefasst

Fachartikel in der Zeitschrift SID Symposium Digest of Technical Papers, online erschienen am 11.09.2026. Verfasst von Liangwei Chen und 1 weiteren Autor:innen. Die Arbeit nennt 14 Literaturangaben.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Computational color constancy remains a pivotal yet ill‐posed problem in computer vision, aiming to recover the canonical appearance of scenes by estimating the unknown scene illuminant. While Convolutional Neural Networks (CNNs) have achieved state‐of‐the‐art performance, existing architectures typically rely on Multi‐Layer Perceptrons (MLPs) for the final regression of high‐dimensional features to illuminant vectors. However, MLPs suffer from inherent limitations in approximating complex non‐linear functions efficiently due to their fixed node‐based activation functions. In this paper, we propose CNN‐KAN, a novel hybrid architecture that integrates the feature extraction power of CNNs with the expressive regression capabilities of Kolmogorov‐Arnold Networks (KANs). Unlike traditional MLPs, KANs utilize learnable univariate B‐spline activation functions on edges, enabling precise modeling of non‐linear color space transformations with high parameter efficiency. We demonstrate that replacing the MLP head with a KAN‐based regressor significantly enhances estimation accuracy. Evaluated on the standard Shi‐Gehler benchmark, our method reduces the mean angular error from 3.42°(CNN‐MLP baseline) to 2.70°, with notable improvements in worst‐case robustness. The proposed framework offers a mathematically elegant and effective alternative for illuminant estimation tasks.

Zitieren

APA

Chen, L., & Luo, M. R. (2026). 6‐2: Beyond MLPs: Convolutional Color Constancy with Kolmogorov‐Arnold Networks. SID Symposium Digest of Technical Papers, 57(S1), 45-48. https://doi.org/10.1002/sdtp.70003_12

Vancouver

Chen L, Luo MR. 6‐2: Beyond MLPs: Convolutional Color Constancy with Kolmogorov‐Arnold Networks. SID Symposium Digest of Technical Papers. 2026;57(S1):45-48. doi:10.1002/sdtp.70003_12

BibTeX
@article{Chen2026,
  author = {Chen, Liangwei and Luo, Ming Ronnier},
  title = {6‐2: Beyond MLPs: Convolutional Color Constancy with Kolmogorov‐Arnold Networks},
  journal = {SID Symposium Digest of Technical Papers},
  year = {2026},
  volume = {57},
  number = {S1},
  pages = {45--48},
  doi = {10.1002/sdtp.70003_12}
}
Kostenlos starten →