Scholarlink

<p>Getting rid of haze was not easy due to color distortion, undesired detail, and the difficulty of working under really foggy conditions. Introduced a novel end-to-end image dehazing framework, which unites multiscale feature enhancement and reinforcement learning into a complete solution. We developed an original bilateral filter multiscale image decomposition to represent a haze-free image while preserving, and exploiting, rich detail information. Image enhancement is an iterative process. The unique feature extraction module with modified smoothing enforces the dynamics of image enhancement that extracts fine details. Channel-spatial attention mechanisms used in the feature fusion module adaptively fuse feature maps proving to be a major step across dynamic scales within the dynamic image enhancement process. To avoid potential extraneous information-induced artifacts, the Vision Transformer fosters haze density assessment with pertinent context references. Last, the SARSA reinforcement training algorithm instructs in performing the optimization iteratively for image enhancement. Experiments carried out on foggy images give evidence that this approach is superior to the existing methods in rendering image details and natural colors while effectively removing haze. The proposed system can deliver significant results in terms of visual quality and be more robust under varying low light and haze conditions.</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 Vision Transformer and Reinforcement Learning-Based Multiscale Framework for Detail-Preserving Image Dehazing

  • 期刊论文

  • 发表日期：

    2025年9月24日

查看全文
  • Yan Cui,
  • Amer Shakir Bin Zainol

摘要

Getting rid of haze was not easy due to color distortion, undesired detail, and the difficulty of working under really foggy conditions. Introduced a novel end-to-end image dehazing framework, which unites multiscale feature enhancement and reinforcement learning into a complete solution. We developed an original bilateral filter multiscale image decomposition to represent a haze-free image while preserving, and exploiting, rich detail information. Image enhancement is an iterative process. The unique feature extraction module with modified smoothing enforces the dynamics of image enhancement that extracts fine details. Channel-spatial attention mechanisms used in the feature fusion module adaptively fuse feature maps proving to be a major step across dynamic scales within the dynamic image enhancement process. To avoid potential extraneous information-induced artifacts, the Vision Transformer fosters haze density assessment with pertinent context references. Last, the SARSA reinforcement training algorithm instructs in performing the optimization iteratively for image enhancement. Experiments carried out on foggy images give evidence that this approach is superior to the existing methods in rendering image details and natural colors while effectively removing haze. The proposed system can deliver significant results in terms of visual quality and be more robust under varying low light and haze conditions.