Scholarlink

<p>Image restoration is a pivotal challenge in computer vision and artificial intelligence, particularly when addressing complex degradations caused by camera motion and dynamic scenes. Traditional methods often struggle to capture degradation features effectively. This paper introduces MCDENet, a novel network architecture that enhances image restoration performance through deep fusion of frequency and spatial domain features. MCDENet comprises four key modules: a multi-scale cross-domain feature extraction module (MSFEM) for capturing local and global context; a cross-domain feature fusion module (CDFFM) for dynamic interaction between spatial and frequency domains; a cross-domain feature enhancement module (CDFEM) for refining feature reconstruction; and a cross-domain feature optimization module (CDFOM) for balancing global semantic consistency and local detail restoration. Systematic evaluations on benchmark datasets demonstrate MCDENet’s competitive performance in deblurring (PSNR of 34.57&#xa0;dB, SSIM of 0.972 on GOPRO) and its effectiveness in deraining and dehazing tasks, with a single trained model directly applied to all three tasks without task-specific retraining, showcasing strong cross-task generalization and adaptability. Experimental results demonstrate that MCDENet achieves a performance improvement of up to 1.5% in PSNR and a 5% reduction in perceptual error over state-of-the-art methods, underscoring its efficacy in enhancing both reconstruction fidelity and visual perceptual quality.</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

Enhanced image restoration via multi-scale cross-domain feature fusion

  • 期刊论文

  • 发表日期：

    2026年3月25日

查看全文
  • Huili Zhang,
  • Yang Zhang,
  • Songhao Zhu,
  • Zhiwei Liang

摘要

Image restoration is a pivotal challenge in computer vision and artificial intelligence, particularly when addressing complex degradations caused by camera motion and dynamic scenes. Traditional methods often struggle to capture degradation features effectively. This paper introduces MCDENet, a novel network architecture that enhances image restoration performance through deep fusion of frequency and spatial domain features. MCDENet comprises four key modules: a multi-scale cross-domain feature extraction module (MSFEM) for capturing local and global context; a cross-domain feature fusion module (CDFFM) for dynamic interaction between spatial and frequency domains; a cross-domain feature enhancement module (CDFEM) for refining feature reconstruction; and a cross-domain feature optimization module (CDFOM) for balancing global semantic consistency and local detail restoration. Systematic evaluations on benchmark datasets demonstrate MCDENet’s competitive performance in deblurring (PSNR of 34.57 dB, SSIM of 0.972 on GOPRO) and its effectiveness in deraining and dehazing tasks, with a single trained model directly applied to all three tasks without task-specific retraining, showcasing strong cross-task generalization and adaptability. Experimental results demonstrate that MCDENet achieves a performance improvement of up to 1.5% in PSNR and a 5% reduction in perceptual error over state-of-the-art methods, underscoring its efficacy in enhancing both reconstruction fidelity and visual perceptual quality.