Asymmetric Frequency-Decoupled Network for Robust Visible–Infrared Fire Detection.
Saved in:
| Title: | Asymmetric Frequency-Decoupled Network for Robust Visible–Infrared Fire Detection. |
|---|---|
| Authors: | Chen, Hongkai1 (AUTHOR), Cai, Hongtu2 (AUTHOR), Sun, Rong1,3 (AUTHOR), Chen, Xiumei1,3 (AUTHOR) chen-xiumei@fzu.edu.cn, Chen, Xin1,2 (AUTHOR), Zhu, Xiaoxing1,3 (AUTHOR) |
| Source: | Remote Sensing. Jun2026, Vol. 18 Issue 11, p1777. 23p. |
| Subjects: | Discrete wavelet transforms, Fire detectors, Forest fire management, Frequency-domain analysis, Artificial neural networks |
| Abstract: | Highlights: What are the main findings? Infrared low-frequency features usually have stronger local contrast in thermal target regions. They can serve as reliability-related cues to guide the adaptive aggregation of visible low-frequency semantics. Smoke occlusion often causes low responses in visible high-frequency features. These low-response features can be converted into an occlusion-related soft gate through negative mapping, which adjusts the complementarity between visible and infrared high-frequency structural information. What are the implications of the main findings? The results show that this asymmetric frequency-decoupled framework effectively improves the accuracy of visible–infrared fire detection. AFDNet provides a more targeted frequency-domain modeling strategy for visible–infrared fire detection under smoke occlusion and cross-modal contamination. Wildfires are highly destructive natural disasters posing serious threats to ecosystems. Visible–infrared fusion is an effective paradigm for robust fire detection in complex scenarios. However, existing spatial-domain fusion methods inevitably suffer from cross-modal contamination: the strong thermal radiation of flames dilutes visible textures, while dense visible smoke suppresses infrared targets. To circumvent this bottleneck, we reveal the frequency-domain asymmetry of fire features and propose an Asymmetric Frequency-Decoupled Network (AFDNet). By shifting from symmetric spatial fusion to asymmetric frequency decoupling, AFDNet effectively isolates modal conflicts, enabling targeted feature enhancement without mutual interference. Specifically, a Modality-Specific Frequency Decoupling (MFD) module first employs the Discrete Wavelet Transform to decompose features into low- and high-frequency sub-bands, breaking spatial entanglement. Subsequently, for low-frequency energy, a Thermal-Guided Low-Frequency Aggregation (TLA) module leverages infrared local contrast as a physical prior to guide fusion, ensuring precise thermal localization while preserving visible scene semantics. For high-frequency details, a Smoke-Masked High-Frequency Restoration (SHR) module maps smoke-induced visible high-frequency collapse into a soft reliability gate. This gate introduces infrared details to compensate for smoke-weakened structural cues. Extensive experiments on the RGBT-3M dataset demonstrate that AFDNet achieves state-of-the-art performance, outperforming the second-best method by 3.37% in mAP, while exhibiting exceptional robustness in complex wildfire environments. [ABSTRACT FROM AUTHOR] |
| Copyright of Remote Sensing is the property of MDPI and its content may not be copied or emailed to multiple sites without the copyright holder's express written permission. Additionally, content may not be used with any artificial intelligence tools or machine learning technologies. However, users may print, download, or email articles for individual use. This abstract may be abridged. No warranty is given about the accuracy of the copy. Users should refer to the original published version of the material for the full abstract. (Copyright applies to all Abstracts.) | |
| Database: | Engineering Source |
|
Full text is not displayed to guests.
Login for full access.
|
|
| Abstract: | Highlights: What are the main findings? Infrared low-frequency features usually have stronger local contrast in thermal target regions. They can serve as reliability-related cues to guide the adaptive aggregation of visible low-frequency semantics. Smoke occlusion often causes low responses in visible high-frequency features. These low-response features can be converted into an occlusion-related soft gate through negative mapping, which adjusts the complementarity between visible and infrared high-frequency structural information. What are the implications of the main findings? The results show that this asymmetric frequency-decoupled framework effectively improves the accuracy of visible–infrared fire detection. AFDNet provides a more targeted frequency-domain modeling strategy for visible–infrared fire detection under smoke occlusion and cross-modal contamination. Wildfires are highly destructive natural disasters posing serious threats to ecosystems. Visible–infrared fusion is an effective paradigm for robust fire detection in complex scenarios. However, existing spatial-domain fusion methods inevitably suffer from cross-modal contamination: the strong thermal radiation of flames dilutes visible textures, while dense visible smoke suppresses infrared targets. To circumvent this bottleneck, we reveal the frequency-domain asymmetry of fire features and propose an Asymmetric Frequency-Decoupled Network (AFDNet). By shifting from symmetric spatial fusion to asymmetric frequency decoupling, AFDNet effectively isolates modal conflicts, enabling targeted feature enhancement without mutual interference. Specifically, a Modality-Specific Frequency Decoupling (MFD) module first employs the Discrete Wavelet Transform to decompose features into low- and high-frequency sub-bands, breaking spatial entanglement. Subsequently, for low-frequency energy, a Thermal-Guided Low-Frequency Aggregation (TLA) module leverages infrared local contrast as a physical prior to guide fusion, ensuring precise thermal localization while preserving visible scene semantics. For high-frequency details, a Smoke-Masked High-Frequency Restoration (SHR) module maps smoke-induced visible high-frequency collapse into a soft reliability gate. This gate introduces infrared details to compensate for smoke-weakened structural cues. Extensive experiments on the RGBT-3M dataset demonstrate that AFDNet achieves state-of-the-art performance, outperforming the second-best method by 3.37% in mAP, while exhibiting exceptional robustness in complex wildfire environments. [ABSTRACT FROM AUTHOR] |
|---|---|
| ISSN: | 20724292 |
| DOI: | 10.3390/rs18111777 |
