Scholarlink

<p>Addressing the challenges of incomplete detail restoration and partial structural information loss during the dehazing process in outdoor scenes, we propose a single-image dehazing algorithm based on cross-scale feature aggregation and structural information guidance. Firstly, an image downsampling module based on cross-scale information aggregation is constructed for the extraction of image detail features. Meanwhile, an M3S module grounded on structural information guidance is established for image dehazing processing, and partial convolutions have been employed to mitigate the interference of redundant features and reduce computational complexity. During the image reconstruction phase, we aggregate and enhance dehazing features across various scales to augment the salience of dehazed image features at different resolutions. The experimental results demonstrate that the proposed dehazing network achieves a SSIM of 0.9911 and a PSNR of 35.94&#xa0;dB on the SOTS dataset. Additionally, on the NH-HAZE dataset, the model attains an SSIM of 0.7207 and a PSNR of 21.12&#xa0;dB. In terms of model’s runtime efficiency, the proposed Tiny model exhibits a computational cost of 10.28G FLOPs and an inference latency of merely 6.32&#xa0;ms, demonstrating its capacity for lightweight deployment.</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

A single image dehazing based on cross-scale feature aggregation and structural information guidance

  • 期刊论文

  • 发表日期：

    2025年12月6日

查看全文
  • Zhenfeng Zhao,
  • Ruiyang Wang,
  • Xinlong Yu,
  • Zihan Zhu,
  • Zhen Liu,
  • Wenbang Fan,
  • Quanli Zhao,
  • Zihang Wu,
  • Kun Meng,
  • Min Wang,
  • Li Yuan

摘要

Addressing the challenges of incomplete detail restoration and partial structural information loss during the dehazing process in outdoor scenes, we propose a single-image dehazing algorithm based on cross-scale feature aggregation and structural information guidance. Firstly, an image downsampling module based on cross-scale information aggregation is constructed for the extraction of image detail features. Meanwhile, an M3S module grounded on structural information guidance is established for image dehazing processing, and partial convolutions have been employed to mitigate the interference of redundant features and reduce computational complexity. During the image reconstruction phase, we aggregate and enhance dehazing features across various scales to augment the salience of dehazed image features at different resolutions. The experimental results demonstrate that the proposed dehazing network achieves a SSIM of 0.9911 and a PSNR of 35.94 dB on the SOTS dataset. Additionally, on the NH-HAZE dataset, the model attains an SSIM of 0.7207 and a PSNR of 21.12 dB. In terms of model’s runtime efficiency, the proposed Tiny model exhibits a computational cost of 10.28G FLOPs and an inference latency of merely 6.32 ms, demonstrating its capacity for lightweight deployment.