Scholarlink

<p>Blind super-resolution (SR) aims to reconstruct high-resolution images from low-resolution inputs with unknown degradations. Traditional methods often struggle with complex and unknown degradations, leading to artifacts and detail distortion. To address this, we propose KADANet, a kernel-aware dual-domain adaptive network, which leverages the complementary strengths of frequency details and spatial structures. KADANet consists of a two-stage architecture: a kernel-aware feature extractor (KAFE) that transfers blur kernel estimation into the low-resolution space, and a frequency spatial adaptive enhancement (FSAE) network. The FSAE network incorporates a dynamic Fourier-Mamba block to integrate global frequency and spatial features, enhancing long-range dependencies. Our method demonstrates improved performance on both synthetic and real-world images, effectively adapting well to various degradation models. Extensive experiments show that KADANet achieves superior reconstruction quality, with PSNR improvements of up to 0.08 dB on the Urban100 dataset for scale factor 4, compared to state-of-the-art methods. The source code is available at <a href="https://github.com/Lucky-lili1/KADANet/tree/main">https://github.com/Lucky-lili1/KADANet/tree/main</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

Kernel-aware dual-domain adaptive network: enhancing blind super-resolution performance

  • 期刊论文

  • 发表日期：

    2025年12月13日

查看全文
  • Keying Li,
  • Yingjiang Li,
  • Yaping Deng,
  • Zibo Wei

摘要

Blind super-resolution (SR) aims to reconstruct high-resolution images from low-resolution inputs with unknown degradations. Traditional methods often struggle with complex and unknown degradations, leading to artifacts and detail distortion. To address this, we propose KADANet, a kernel-aware dual-domain adaptive network, which leverages the complementary strengths of frequency details and spatial structures. KADANet consists of a two-stage architecture: a kernel-aware feature extractor (KAFE) that transfers blur kernel estimation into the low-resolution space, and a frequency spatial adaptive enhancement (FSAE) network. The FSAE network incorporates a dynamic Fourier-Mamba block to integrate global frequency and spatial features, enhancing long-range dependencies. Our method demonstrates improved performance on both synthetic and real-world images, effectively adapting well to various degradation models. Extensive experiments show that KADANet achieves superior reconstruction quality, with PSNR improvements of up to 0.08 dB on the Urban100 dataset for scale factor 4, compared to state-of-the-art methods. The source code is available at https://github.com/Lucky-lili1/KADANet/tree/main.