Scholarlink

<p>Atmospheric particles such as dust, smoke, and other pollutants degrade the quality of visual content by scattering light, leading to reduced contrast and the appearance of a white veil commonly referred to as haze. To address this issue, various dehazing techniques have been developed to restore clarity and enhance the visual appeal of affected images. This study proposes an advanced dehazing framework, termed enhanced DehazeFormer, which integrates a comprehensive pre-processing pipeline, a transformer-based dehazing model, and a structure-aware filtering mechanism to produce high-quality results. The pre-processing stage involves two key steps aimed at correcting measurement-related distortions and enhancing visual features. Initially, a homomorphic filtering technique is applied to manage dynamic range and illumination inconsistencies. This is followed by Contrast Limited Adaptive Histogram Equalization (CLAHE), which improves local contrast prior to the main dehazing phase. The core of the method, DehazeFormer, builds upon the Swin Transformer architecture, adapting it specifically for haze removal tasks. To further refine the transmission map, a structure-guided filter (Sℓo) is utilized, effectively eliminating artifacts while preserving the global structure of the guidance image. The resulting images and frames demonstrate notable visual improvement. To rigorously evaluate performance, extensive experiments were conducted on visible light images, Near Infrared (NIR) frames, and real-world hazy images. A variety of objective metrics, including correlation, Peak Signal-to-Noise Ratio (PSNR), Feature Similarity Index (FSIM), Feature Similarity Index with Chrominance (FSIMC), Structural Similarity Index (SSIM), and entropy, were employed to quantify image quality. Additionally, histogram analysis and spectral entropy were used to benchmark the proposed method against existing approaches. Experimental findings affirm the superiority of the enhanced DehazeFormer, with substantial improvements in both visual and numerical performance. Specifically, average PSNR gains across five frames show enhancement percentages of 92.99% for visible images and 117.61% for NIR frames, underscoring the robustness and effectiveness of the proposed solution.</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

DehazeFormer: a swintransformer-based approach for high-quality haze removal in images and videos

  • 期刊论文

  • 发表日期：

    2026年2月9日

查看全文
  • Abeer Ayoub,
  • Walid El-Shafai,
  • Mohamed Aouf,
  • Fathi E. Abd El-Samie,
  • Ehab K. I. Hamad,
  • S. EL-Rabaie

摘要

Atmospheric particles such as dust, smoke, and other pollutants degrade the quality of visual content by scattering light, leading to reduced contrast and the appearance of a white veil commonly referred to as haze. To address this issue, various dehazing techniques have been developed to restore clarity and enhance the visual appeal of affected images. This study proposes an advanced dehazing framework, termed enhanced DehazeFormer, which integrates a comprehensive pre-processing pipeline, a transformer-based dehazing model, and a structure-aware filtering mechanism to produce high-quality results. The pre-processing stage involves two key steps aimed at correcting measurement-related distortions and enhancing visual features. Initially, a homomorphic filtering technique is applied to manage dynamic range and illumination inconsistencies. This is followed by Contrast Limited Adaptive Histogram Equalization (CLAHE), which improves local contrast prior to the main dehazing phase. The core of the method, DehazeFormer, builds upon the Swin Transformer architecture, adapting it specifically for haze removal tasks. To further refine the transmission map, a structure-guided filter (Sℓo) is utilized, effectively eliminating artifacts while preserving the global structure of the guidance image. The resulting images and frames demonstrate notable visual improvement. To rigorously evaluate performance, extensive experiments were conducted on visible light images, Near Infrared (NIR) frames, and real-world hazy images. A variety of objective metrics, including correlation, Peak Signal-to-Noise Ratio (PSNR), Feature Similarity Index (FSIM), Feature Similarity Index with Chrominance (FSIMC), Structural Similarity Index (SSIM), and entropy, were employed to quantify image quality. Additionally, histogram analysis and spectral entropy were used to benchmark the proposed method against existing approaches. Experimental findings affirm the superiority of the enhanced DehazeFormer, with substantial improvements in both visual and numerical performance. Specifically, average PSNR gains across five frames show enhancement percentages of 92.99% for visible images and 117.61% for NIR frames, underscoring the robustness and effectiveness of the proposed solution.