Scholarlink

<p>Although image deblurring methods have achieved impressive results in terms of PSNR in recent years, most existing approaches still suffer from high GPU memory consumption. To address this issue, we develop an approach to reduce GPU memory consumption. Specifically, the conventional encoder-decoder architecture of the Transformer network is restructured into a reversible architecture, reducing memory consumption during backpropagation. Additionally, a Multi-Scale Feature Fusion Module (MSFM) is introduced between the encoder and decoder to fuse low-level details and high-level semantic information. To handle the spatial variability of blur kernels across images, a nonlinear feature processing mechanism is presented, improving blur removal at varying levels. Furthermore, a Channel Importance Learner (CIL) mechanism is designed, which prioritizes the processing of high-value feature channels, thereby enhancing the network’s learning efficiency and overall performance. Experimental results demonstrate that, when processing <InlineEquation ID="IEq1"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="11760_2025_4757_Article_IEq1.gif" Format="GIF" Height="14" Rendition="HTML" Resolution="72" Type="Linedraw" Width="70" /> </InlineMediaObject> <EquationSource Format="TEX">\(256 \times 256\)</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mn>256</mn> <mo>×</mo> <mn>256</mn> </mrow> </math></EquationSource> </InlineEquation> images, the memory consumption is 3395 MB, representing at least a 28.13% reduction compared to existing methods.</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

Reversible nonlinear transformer for image deblurring

  • 期刊论文

  • 发表日期：

    2025年9月22日

查看全文
  • Guanyuan Feng,
  • Xin Zhang,
  • Weili Shi,
  • Yu Miao

摘要

Although image deblurring methods have achieved impressive results in terms of PSNR in recent years, most existing approaches still suffer from high GPU memory consumption. To address this issue, we develop an approach to reduce GPU memory consumption. Specifically, the conventional encoder-decoder architecture of the Transformer network is restructured into a reversible architecture, reducing memory consumption during backpropagation. Additionally, a Multi-Scale Feature Fusion Module (MSFM) is introduced between the encoder and decoder to fuse low-level details and high-level semantic information. To handle the spatial variability of blur kernels across images, a nonlinear feature processing mechanism is presented, improving blur removal at varying levels. Furthermore, a Channel Importance Learner (CIL) mechanism is designed, which prioritizes the processing of high-value feature channels, thereby enhancing the network’s learning efficiency and overall performance. Experimental results demonstrate that, when processing \(256 \times 256\) 256 × 256 images, the memory consumption is 3395 MB, representing at least a 28.13% reduction compared to existing methods.