Density-Guided and Frequency Modulation Dehazing Network for Remote Sensing Images
Remote sensing image (RSI) dehazing methods have gained significant attention for their ability to restore clear images, which are crucial for applications such as mineral exploration and flood range forecasting. The haze present in foggy RSIs is typically nonhomogeneous, posing a challenge for exis...
Saved in:
| Main Authors: | , , , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
IEEE
2025-01-01
|
| Series: | IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing |
| Subjects: | |
| Online Access: | https://ieeexplore.ieee.org/document/10946836/ |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Remote sensing image (RSI) dehazing methods have gained significant attention for their ability to restore clear images, which are crucial for applications such as mineral exploration and flood range forecasting. The haze present in foggy RSIs is typically nonhomogeneous, posing a challenge for existing dehazing methods, which often struggle with the effective removal of such haze. Furthermore, current approaches primarily address foggy inputs within the spatial domain, neglecting the potential advantages of exploring the frequency domain for dehazing. To address these challenges, in this article, we propose a density-guided and frequency modulation dehazing network (DFDNet) specifically designed for RSI dehazing. The DFDNet integrates a density-guided transformer dehazing subnet (DTDN) and a frequency dual-path enhancing subnet (FDEN), enabling the restoration of clear RSIs by combined spatial- and frequency-domain processing. The DTDN leverages the dark channel prior to generate the density-aware attention and guide the removal of nonhomogeneous haze within the spatial domain. The FDEN is a dual-path structure that modulates the frequencies to enhance the details of dehazed images produced by the DTDN. Comprehensive quantitative and qualitative evaluations on StateHaze1k, RICE, and RRSD300 datasets demonstrate the superiority and generalization of the proposed DFDNet. Especially, the proposed DFDNet outperforms recent FSNet by 0.81 dB on the StateHaze1k-thick dataset. |
|---|---|
| ISSN: | 1939-1404 2151-1535 |