Scholarlink

<p>In recent years, significant progress has been made in learning-based blind single-image super-resolution (SISR) methods, where the accurate estimation of the blur kernel is a key factor influencing reconstruction quality. However, due to the lack of real paired data and the complexity of kernel prior modeling, blur kernel estimation still faces challenges such as insufficient precision and poor generalization ability. To address this, this paper proposes a kernel discriminative mechanism-based unsupervised blind super-resolution method (KDSR), which enhances the robustness of blur kernel estimation by jointly optimizing the kernel estimation module and the kernel discriminator through an adversarial learning strategy. For image high-frequency detail restoration, this paper designs a dynamic multi-scale super-resolution network (DMNet), which achieves global context awareness and local structure enhancement through multi-scale attention and channel-space dual-branch attention modules, generating multi-granularity feature representations and effectively addressing potential issues with local optima in the reconstruction process. Furthermore, the proposed method combines global semantic guidance and local texture constraints for kernel feature estimation, effectively alleviating kernel local optimum problems. Extensive experimental results on the Set5, Set14, BSD100, and Urban100 datasets show that the proposed KDSR method produces higher-quality super-resolved images compared to existing methods, while also demonstrating strong generalization ability.</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

Multi-scale attention and kernel discrimination for image blind super-resolution

  • 期刊论文

  • 发表日期：

    2026年5月19日

查看全文
  • Zhicheng Wang,
  • Jinhu Wu,
  • Longge Wang

摘要

In recent years, significant progress has been made in learning-based blind single-image super-resolution (SISR) methods, where the accurate estimation of the blur kernel is a key factor influencing reconstruction quality. However, due to the lack of real paired data and the complexity of kernel prior modeling, blur kernel estimation still faces challenges such as insufficient precision and poor generalization ability. To address this, this paper proposes a kernel discriminative mechanism-based unsupervised blind super-resolution method (KDSR), which enhances the robustness of blur kernel estimation by jointly optimizing the kernel estimation module and the kernel discriminator through an adversarial learning strategy. For image high-frequency detail restoration, this paper designs a dynamic multi-scale super-resolution network (DMNet), which achieves global context awareness and local structure enhancement through multi-scale attention and channel-space dual-branch attention modules, generating multi-granularity feature representations and effectively addressing potential issues with local optima in the reconstruction process. Furthermore, the proposed method combines global semantic guidance and local texture constraints for kernel feature estimation, effectively alleviating kernel local optimum problems. Extensive experimental results on the Set5, Set14, BSD100, and Urban100 datasets show that the proposed KDSR method produces higher-quality super-resolved images compared to existing methods, while also demonstrating strong generalization ability.