Scholarlink

<p>Video dehazing aims to restore clear and high-contrast video frames from hazy inputs. However, the non-homogeneous spatial distribution of haze and its temporal variations pose significant challenges. To address these issues, we propose a novel adaptive spatiotemporal partitioning video dehazing network (ASTP-Net). Specifically, a sliding window correlation attention (SWCA) module is proposed to capture interactive haze density information locally and establish long-range spatial dependencies. In addition, we propose a temporal neighborhood complementary attention (TNCA) module to adaptively aggregate complementary information between reference frames, mitigating temporal variations. Comprehensive experiments demonstrate that our ASTP-Net outperforms state-of-the-art video dehazing methods, achieving a 2.43&#xa0;dB improvement in PSNR and a 2.38% improvement in SSIM on benchmark datasets. The proposed approach not only enhances dehazing performance, but also demonstrates improved computational efficiency. The dataset, source code, and usage guidelines are available at the following link: <a href="https://github.com/CodeStarting-design/ASTP-net">https://github.com/CodeStarting-design/ASTP-net</a>.</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

Adaptive spatiotemporal partitioning for efficient video dehazing

  • 期刊论文

  • 发表日期：

    2025年8月21日

查看全文
  • Wang Zhen,
  • Liu Yanli,
  • Xing Guanyu,
  • Wei Housheng

摘要

Video dehazing aims to restore clear and high-contrast video frames from hazy inputs. However, the non-homogeneous spatial distribution of haze and its temporal variations pose significant challenges. To address these issues, we propose a novel adaptive spatiotemporal partitioning video dehazing network (ASTP-Net). Specifically, a sliding window correlation attention (SWCA) module is proposed to capture interactive haze density information locally and establish long-range spatial dependencies. In addition, we propose a temporal neighborhood complementary attention (TNCA) module to adaptively aggregate complementary information between reference frames, mitigating temporal variations. Comprehensive experiments demonstrate that our ASTP-Net outperforms state-of-the-art video dehazing methods, achieving a 2.43 dB improvement in PSNR and a 2.38% improvement in SSIM on benchmark datasets. The proposed approach not only enhances dehazing performance, but also demonstrates improved computational efficiency. The dataset, source code, and usage guidelines are available at the following link: https://github.com/CodeStarting-design/ASTP-net.