.png){kind=icon}
.png){kind=icon}
Return
SFSNet: object detection in adverse weather via active frequency-spatial feature recovery
.png){kind=icon}
PRE
AI
Z
Y
DOI：10.1007/s00530-026-02593-3.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Object detection in adverse weather remains challenging as visibility degradation severely impairs the feature extraction capabilities of visual recognition systems. Existing “Restore-then-Detect” paradigms usually optimize image restoration for human visual quality. These methods may smooth high-frequency structures that are essential for object localization and recognition. To tackle this problem, we propose SFSNet, which performs real-time frequency-spatial feature recovery for object detection under hazy and low-light conditions. SFSNet consists of three main modules. First, the Adaptive Spatial-Wavelet Illumination Estimator (AS-WIE) performs spatial-adaptive amplitude boosting on the low-frequency sub-band for robust illumination recovery. The recovery process is supervised by a Physical Illumination Loss (PIL) that constrains the illumination estimator to align with physical illumination distributions. PIL prevents over-amplification and enhances photometric consistency. Second, the Quad-Prior Aggregation (QPA) module reformulates physical priors into learnable feature-level constraints to distinguish degradation from semantic content. Third, a Symmetric Frequency-Spatial Architecture is proposed to replace standard pooling with invertible Discrete Wavelet Transform (DWT) for information-preserving decomposition. The symmetric frequency-spatial architecture employs the Structure Sharpening (SS) method to recover high-frequency structural cues while suppressing sensor noise. Experimental results show that SFSNet+YOLOv8s achieves state-of-the-art performance on RTTS (56.8% mAP) and ExDark (58.8% mAP). Ablation studies and visualizations also demonstrate that SFSNet improves object detection in adverse weather.
Keywords:
Object detection
Adverse weather
Feature recovery
Frequency-spatial analysis
Physical priors
Journal
IF:
3.1
Papers:
2.7K
Citations:
2.7K
Organization
Cited Papers