Scholarlink

<p>In practical applications, image resolution has already reached 4K. However, large images often contain many smooth regions that can achieve good performance using networks with fewer channels. Existing methods for accelerating SR models typically divide images into multiple patches and process them through separate branches. However, these approaches suffer from two limitations: (1) the lack of scalability to be developed on platforms with different capacities and (2) the lack of interaction between multiple branches results in limited acceleration performance. Based on this, we propose adaptive branch selection (ABS) for accelerating image super-resolution. ABS utilizes an efficient regressor to predict the performance increment between branches, dynamically selecting different branches for each patch by adjusting thresholds. To further enhance the acceleration performance, we introduce Progressive Mutual Information Knowledge Distillation (PMID) to help improve the SR performance of branches with fewer channels. Experimental results on the Test8K dataset show that FSRCNN-ABS achieves performance comparable to the original model while using only 49<InlineEquation ID="IEq1"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="371_2025_4034_Article_IEq1.gif" Format="GIF" Height="16" Rendition="HTML" Resolution="72" Type="Linedraw" Width="15" /> </InlineMediaObject> <EquationSource Format="TEX">\(\%\)</EquationSource> <EquationSource Format="MATHML"><math> <mo>%</mo> </math></EquationSource> </InlineEquation> of the FLOPs. The codes are available at: <a href="https://github.com/dcgithubtools/ABS.git">https://github.com/dcgithubtools/ABS.git</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 branch selection for accelerate image super-resolution

  • 期刊论文

  • 发表日期：

    2025年6月25日

查看全文
  • Cheng Ding,
  • Zhongqiu Zhao,
  • Hao Shen,
  • Xiufeng Liu

摘要

In practical applications, image resolution has already reached 4K. However, large images often contain many smooth regions that can achieve good performance using networks with fewer channels. Existing methods for accelerating SR models typically divide images into multiple patches and process them through separate branches. However, these approaches suffer from two limitations: (1) the lack of scalability to be developed on platforms with different capacities and (2) the lack of interaction between multiple branches results in limited acceleration performance. Based on this, we propose adaptive branch selection (ABS) for accelerating image super-resolution. ABS utilizes an efficient regressor to predict the performance increment between branches, dynamically selecting different branches for each patch by adjusting thresholds. To further enhance the acceleration performance, we introduce Progressive Mutual Information Knowledge Distillation (PMID) to help improve the SR performance of branches with fewer channels. Experimental results on the Test8K dataset show that FSRCNN-ABS achieves performance comparable to the original model while using only 49 \(\%\) % of the FLOPs. The codes are available at: https://github.com/dcgithubtools/ABS.git.