Scholarlink

<p>Object detection under foggy weather is hindered by frequency-selective atmospheric scattering and spatially non-uniform fog distribution, while dehazing-then-detection pipelines additionally suffer from error accumulation. We present a feature-aware framework built upon YOLOv11 that addresses these degradations through a hierarchical enhancement pipeline along the feature pyramid. At the shallow stage, the High-Low Frequency Decoupling (HLFD) module restores scattering-eroded edges and textures by enhancing high- and low-frequency components separately. At the deep bottleneck, the Spatial Pyramid Pooling Fast with Cross Stage Partial Connection (SPPFCSPC) module aggregates global context through serial cascaded pooling at reduced cost. At the cross-scale neck, the Content-Adaptive Feature Modulation Fusion (CAFMFusion) module replaces fixed-weight fusion with dual-level content-adaptive gating that reweights shallow and deep features according to local fog density. On the real-world RTTS benchmark, the framework attains 60.4% mAP at 95.5&#xa0;FPS; cross-dataset evaluation on Foggy Cityscapes and cross-backbone transfer to YOLOv8/v9/v12 confirm that the gains are not tied to a single architecture. The compact, real-time footprint suits multi-stream concurrent inference on GPU-accelerated and HPC infrastructure for city-scale foggy perception.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Enhancing foggy weather object detection via frequency decoupling, content-adaptive feature fusion and efficient spatial pyramid pooling in YOLOv11

  • 期刊论文

  • 发表日期：

    2026年7月30日

查看全文
  • Jiaxin Zhu

摘要

Object detection under foggy weather is hindered by frequency-selective atmospheric scattering and spatially non-uniform fog distribution, while dehazing-then-detection pipelines additionally suffer from error accumulation. We present a feature-aware framework built upon YOLOv11 that addresses these degradations through a hierarchical enhancement pipeline along the feature pyramid. At the shallow stage, the High-Low Frequency Decoupling (HLFD) module restores scattering-eroded edges and textures by enhancing high- and low-frequency components separately. At the deep bottleneck, the Spatial Pyramid Pooling Fast with Cross Stage Partial Connection (SPPFCSPC) module aggregates global context through serial cascaded pooling at reduced cost. At the cross-scale neck, the Content-Adaptive Feature Modulation Fusion (CAFMFusion) module replaces fixed-weight fusion with dual-level content-adaptive gating that reweights shallow and deep features according to local fog density. On the real-world RTTS benchmark, the framework attains 60.4% mAP at 95.5 FPS; cross-dataset evaluation on Foggy Cityscapes and cross-backbone transfer to YOLOv8/v9/v12 confirm that the gains are not tied to a single architecture. The compact, real-time footprint suits multi-stream concurrent inference on GPU-accelerated and HPC infrastructure for city-scale foggy perception.