Scholarlink

<p>The field of image restoration focuses on recovering high-quality visual content from degraded inputs. In practical applications, images often encounter complex and varied forms of degradation, posing significant challenges for existing restoration models. To address this, we propose AM-PromptIR, an adaptive multi-modal image restoration framework that leverages implicit correlations between degradation types to guide the restoration process. Specifically, we propose an Adaptive Multi-Modal Prompt Generation that leverages a language-image model to capture the implicit relationships between various degradation types. Additionally, we introduce Simplified Mixed Attention Mechanism and Feature Enhancement and Fusion Module to improve feature extraction and fusion, respectively. By 9.2% in dehazing tasks did AM-PromptIR improve PSNR compared to the PromptIR baseline model; by 5.3% in deraining operations was the PSNR enhancement; and across various degraded tasks, an average increase of 3.0% in PSNR was achieved. These findings highlight the robustness and effectiveness of our approach in handling diverse image degradations. Our code is available here: <a href="https://github.com/buuzhangsen/AM-PromptIR">https://github.com/buuzhangsen/AM-PromptIR</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 multi-modal prompting for universal image restoration amidst diverse degradations

  • 期刊论文

  • 发表日期：

    2025年7月2日

查看全文
  • Sen Zhang,
  • Jinhua Wang,
  • Ning He,
  • Sunhan Xu,
  • Shuai Liu,
  • Pengcheng Yu,
  • Xiaoyue Ma

摘要

The field of image restoration focuses on recovering high-quality visual content from degraded inputs. In practical applications, images often encounter complex and varied forms of degradation, posing significant challenges for existing restoration models. To address this, we propose AM-PromptIR, an adaptive multi-modal image restoration framework that leverages implicit correlations between degradation types to guide the restoration process. Specifically, we propose an Adaptive Multi-Modal Prompt Generation that leverages a language-image model to capture the implicit relationships between various degradation types. Additionally, we introduce Simplified Mixed Attention Mechanism and Feature Enhancement and Fusion Module to improve feature extraction and fusion, respectively. By 9.2% in dehazing tasks did AM-PromptIR improve PSNR compared to the PromptIR baseline model; by 5.3% in deraining operations was the PSNR enhancement; and across various degraded tasks, an average increase of 3.0% in PSNR was achieved. These findings highlight the robustness and effectiveness of our approach in handling diverse image degradations. Our code is available here: https://github.com/buuzhangsen/AM-PromptIR.