Scholarlink

<p>Video dehazing is a complex challenge, particularly in dynamic scenes affected by adverse weather conditions. Heterogeneous haze densities complicate haze removal, and interpreting scene depth in real time is difficult due to varying depth complexities. The lack of a standardized benchmark dataset for hazy videos further highlights the need for scalable solutions. This work introduces a hybrid approach to address these issues, comprising four key phases: airlight estimation, transmission map estimation, deep visual priors generation, and contrast refinement. Airlight is estimated using a box filter applied to the minimum channel fusion matrix. Transmission estimation captures dynamic haze densities by applying a Fast Guided Filter to refined RGB channels, while scene depth information is extracted using sigmoid features. The resultant matrix is then processed by a pre-trained InceptionNet-V3 to generate scene-specific deep visual priors. Finally, CLAHE-based contrast enhancement refines the output, producing dehazed frames. The method is evaluated on subsets of RESIDE, URHI, and self-collected PanopticVisionHaze datasets, demonstrating superior performance over state-of-the-art dehazing methods.</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

Deep visualization prior based video dehazing: a hybrid approach for enhanced visibility

  • 期刊论文

  • 发表日期：

    2025年6月8日

查看全文
  • Yasmeen Khaliq,
  • Aun Irtaza,
  • Syed Muhammad Adnan Shah

摘要

Video dehazing is a complex challenge, particularly in dynamic scenes affected by adverse weather conditions. Heterogeneous haze densities complicate haze removal, and interpreting scene depth in real time is difficult due to varying depth complexities. The lack of a standardized benchmark dataset for hazy videos further highlights the need for scalable solutions. This work introduces a hybrid approach to address these issues, comprising four key phases: airlight estimation, transmission map estimation, deep visual priors generation, and contrast refinement. Airlight is estimated using a box filter applied to the minimum channel fusion matrix. Transmission estimation captures dynamic haze densities by applying a Fast Guided Filter to refined RGB channels, while scene depth information is extracted using sigmoid features. The resultant matrix is then processed by a pre-trained InceptionNet-V3 to generate scene-specific deep visual priors. Finally, CLAHE-based contrast enhancement refines the output, producing dehazed frames. The method is evaluated on subsets of RESIDE, URHI, and self-collected PanopticVisionHaze datasets, demonstrating superior performance over state-of-the-art dehazing methods.