SFFDM: Diffusion model-based spatial-frequency fusion network for remote

Sun, Haoqin

doi:10.1016/j.image.2026.117663

Return

SFFDM: Diffusion model-based spatial-frequency fusion network for remote

2026-07-01

PRE

DOI：10.1016/j.image.2026.117663

Original

Original request for help

Save

Abstract

En 中文

In recent years, deep neural networks have been widely applied to remote sensing image dehazing and have achieved remarkable progress. However, features extracted from the spatial and frequency domains are considered to have different representation focuses and carry distinct types of information, which exhibit a certain degree of complementarity. This complementarity is believed to be beneficial for learning more representative and robust feature representations. Furthermore, how to effectively extract noise information in the frequency domain and utilize it to guide more accurate denoising remains one of the key challenges in the field of image dehazing. Based on this, a Diffusion Model-based Spatial-Frequency Fusion network (SFFDM) is proposed for remote sensing image dehazing. Specifically, a Haze Information Extraction Module (HIEM) and a Frequency Information Purification Module (FIPM) are designed. The HIEM captures haze-related features from the original haze image in the frequency domain, which are subsequently utilized by the FIPM to enhance dehazing performance. Additionally, the Spatial-Frequency Fusion Module (SFFM) is introduced to integrate features from both domains, contributing to improved image reconstruction quality. Experimental results on multiple datasets demonstrate that the SFFDM network significantly enhances dehazing performance while preserving rich color and texture details.

Keywords:

Deep learning

Diffusion model

Remote sensing image dehazing

Spatial domain

Frequency domain