Scholarlink

<p>Single-image dehazing remains challenging for computer vision applications due to the complex degradation process of haze. Traditional physics-based methods lack adaptability to diverse real-world conditions, while purely data-driven approaches often fail to generalize by neglecting physical constraints. We propose a Physics-Guided Atmospheric Restoration Network (PGARN), a novel architecture that synergistically combines physical principles with deep learning. PGARN features: (1) Multi-path Adaptive Attention Blocks (MAAB) that efficiently extract multi-scale features through parallel convolution paths with varying receptive fields, (2) a Physics-Guided Module (PGM) that explicitly estimates transmissions and atmospheric light parameters, and (3) a Physics Scattering Transformer Block (PSTB) that incorporates these physical parameters into self-attention mechanisms. This integration enables PGARN to process haze at different depths adaptively while preserving local details and global consistency. Comprehensive experiments on RESIDE, Haze4K and RTTS benchmarks demonstrate that PGARN outperforms state-of-the-art methods in objective metrics and visual quality. Furthermore, PGARN substantially enhances downstream task performance, significantly improving object detection accuracy in challenging, hazy conditions.</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

Physics-guided atmospheric restoration network for single image dehazing

  • 期刊论文

  • 发表日期：

    2025年6月24日

查看全文
  • Jingchao Hou,
  • Garas Gendy,
  • Guanghui He

摘要

Single-image dehazing remains challenging for computer vision applications due to the complex degradation process of haze. Traditional physics-based methods lack adaptability to diverse real-world conditions, while purely data-driven approaches often fail to generalize by neglecting physical constraints. We propose a Physics-Guided Atmospheric Restoration Network (PGARN), a novel architecture that synergistically combines physical principles with deep learning. PGARN features: (1) Multi-path Adaptive Attention Blocks (MAAB) that efficiently extract multi-scale features through parallel convolution paths with varying receptive fields, (2) a Physics-Guided Module (PGM) that explicitly estimates transmissions and atmospheric light parameters, and (3) a Physics Scattering Transformer Block (PSTB) that incorporates these physical parameters into self-attention mechanisms. This integration enables PGARN to process haze at different depths adaptively while preserving local details and global consistency. Comprehensive experiments on RESIDE, Haze4K and RTTS benchmarks demonstrate that PGARN outperforms state-of-the-art methods in objective metrics and visual quality. Furthermore, PGARN substantially enhances downstream task performance, significantly improving object detection accuracy in challenging, hazy conditions.