Small-Object Detection in Foggy Scenes via High-Frequency-Enhanced Feature Fusion and Compact State-Space Modeling

In adverse weather and rapidly changing scenes, fog severely reduces image contrast and obscures target textures, causing small-object detection to suffer from feature weakening and background interference. Many existing detectors, meanwhile, rely on computationally intensive feature modeling, making it difficult to achieve real-time inference while effectively mitigating fog-induced degradation. To address these challenges, we propose HS-MambaDet, a frequency-compensated hidden-state state-space detection network for accurate and efficient small-object detection in foggy environments. Specifically, we embed a lightweight SSD-based state-space modeling module with frequency-domain window attention (FWS-SSD) into the backbone, preserving long-range dependency modeling with low computational overhead while emphasizing informative high-frequency details and attenuating low-frequency haze interference. This study highlights a symmetry-inspired balance between global context modeling and local detail restoration. In the neck network, a multi-scale frequency-spatial fusion (MFSF) module further strengthens fine-grained object representations and cross-scale contextual interactions. In addition, we introduce a fog-aware detection loss to better supervise low-contrast and detail-deficient regions, improving detection robustness in foggy scenes. Extensive experiments on RTTS and Cityscapes demonstrate clear and consistent gains: HS-MambaDet outperforms representative one-stage, two-stage, and state-space-based detectors by up to 4.3% in mAP@0.5 and 6.5% in mAP@0.5:0.95, while maintaining competitive inference efficiency, thereby achieving a favorable accuracy-efficiency trade-off for foggy small-object detection.