Physics-Infused Dual-Path Transformer for Single Image Dehazing
IEEE Signal Processing Letters · 2026 · 33 · 1681-1685
DOI: 10.1109/lsp.2026.3682992 ↗Kurz gefasst
Fachartikel in der Zeitschrift IEEE Signal Processing Letters, 2026. Verfasst von Hao Yang und 4 weiteren Autor:innen. Die Arbeit nennt 38 Literaturangaben. Gefördert von: Major Technology Innovation Project of Hubei Province.
Abstract (Originalsprache)
Existing physics-informed dehazing networks typically adopt a loose coupling paradigm, imposing physical constraints primarily through objective functions or input priors, which hinders the explicit integration of physical laws into feature evolution. To address this limitation, we propose a Physics-Infused Dual-Path Transformer (PIDP-Former) that fundamentally redefines the interaction between physical laws and deep features by embedding the Atmospheric Scattering Model (ASM) directly into the network topology. Unlike previous works, PIDP-Former synergizes a physics-aware parameter estimation path with a data-driven feature restoration path through two novel mechanisms: Physics-Guided Spatial Modulation (PGSM) and Multi-scale Physical Residual Feedback (MPRF). This design establishes a latent “perceive-and-correct” loop for iterative feature refinement. Experimental results on SOTS-Outdoor reveal 0.67% PSNR and 0.15% SSIM improvements over state-of-the-art (SOTA) benchmarks, while real-world tests further validate the framework’s robust generalization and physical interpretability.
Förderung
- Major Technology Innovation Project of Hubei Province
Zitieren
APA
Yang, H., Wei, J., Wang, M., Zhu, L., & Xu, G. (2026). Physics-Infused Dual-Path Transformer for Single Image Dehazing. IEEE Signal Processing Letters, 33, 1681-1685. https://doi.org/10.1109/lsp.2026.3682992
Vancouver
Yang H, Wei J, Wang M, Zhu L, Xu G. Physics-Infused Dual-Path Transformer for Single Image Dehazing. IEEE Signal Processing Letters. 2026;33:1681-1685. doi:10.1109/lsp.2026.3682992
BibTeX
@article{Yang2026,
author = {Yang, Hao and Wei, Jia and Wang, Min and Zhu, Lanyan and Xu, Gang},
title = {Physics-Infused Dual-Path Transformer for Single Image Dehazing},
journal = {IEEE Signal Processing Letters},
year = {2026},
volume = {33},
pages = {1681--1685},
doi = {10.1109/lsp.2026.3682992}
}
EUVIMED