Scholarlink

<p>Video dehazing is an essential pre-processing task for improving visibility and reliability in videos captured under atmospheric haze, fog, or underwater scattering. Existing methods often rely on handcrafted priors or large amounts of paired training data, limiting their robustness, temporal consistency, and suitability for real-time deployment. To address these challenges, we propose DINO-Dehaze, a self-supervised and real-time video dehazing framework that jointly models semantic awareness, temporal coherence, and computational efficiency. DINO-Dehaze leverages a DINO-ViT backbone to extract haze-invariant semantic features without requiring ground-truth supervision, enabling strong generalization across diverse environments. These features are refined using an enhanced UNet + architecture equipped with temporal attention and multi-scale fusion, ensuring spatial detail preservation and inter-frame consistency. Furthermore, a probe-based knowledge distillation strategy transfers multi-level representations from a high-capacity teacher model to a lightweight student network, enabling efficient deployment on resource-constrained devices. Extensive experiments on REVIDE, HazeWorld, and D-Hazy datasets demonstrate that DINO-Dehaze consistently outperforms state-of-the-art methods in both real-world and synthetic scenarios, achieving up to 32.6&#xa0;dB PSNR, 0.962 SSIM, and 0.112 LPIPS, while maintaining a compact model size of 18.3&#xa0;MB and real-time inference at 57.2 FPS on an NVIDIA A100 GPU. These results confirm the effectiveness and practicality of the proposed framework for real-time video dehazing applications.</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

DINO-Dehaze: Self-Supervised Transformer-Guided Video Dehazing with UNet + and Probe Distillation

  • 期刊论文

  • 发表日期：

    2026年4月21日

查看全文
  • Sandeep Vishwakarma,
  • Deepika Punj,
  • Anuradha Pillai

摘要

Video dehazing is an essential pre-processing task for improving visibility and reliability in videos captured under atmospheric haze, fog, or underwater scattering. Existing methods often rely on handcrafted priors or large amounts of paired training data, limiting their robustness, temporal consistency, and suitability for real-time deployment. To address these challenges, we propose DINO-Dehaze, a self-supervised and real-time video dehazing framework that jointly models semantic awareness, temporal coherence, and computational efficiency. DINO-Dehaze leverages a DINO-ViT backbone to extract haze-invariant semantic features without requiring ground-truth supervision, enabling strong generalization across diverse environments. These features are refined using an enhanced UNet + architecture equipped with temporal attention and multi-scale fusion, ensuring spatial detail preservation and inter-frame consistency. Furthermore, a probe-based knowledge distillation strategy transfers multi-level representations from a high-capacity teacher model to a lightweight student network, enabling efficient deployment on resource-constrained devices. Extensive experiments on REVIDE, HazeWorld, and D-Hazy datasets demonstrate that DINO-Dehaze consistently outperforms state-of-the-art methods in both real-world and synthetic scenarios, achieving up to 32.6 dB PSNR, 0.962 SSIM, and 0.112 LPIPS, while maintaining a compact model size of 18.3 MB and real-time inference at 57.2 FPS on an NVIDIA A100 GPU. These results confirm the effectiveness and practicality of the proposed framework for real-time video dehazing applications.