Scholarlink

<p>Image restoration aims to recover high-quality visual content from degraded inputs, playing a pivotal role in various real-world applications. While Transformer-based architectures have achieved state-of-the-art performance due to their robust global modeling capabilities, their deployment is severely hindered by inherent quadratic computational complexity. To address this challenge, we propose an efficient aligned kernel network (AKNet). At its core lies the aligned kernel module (AKModule), which innovatively employs super-large convolution kernels to capture global receptive fields with minimal computational overhead, effectively mimicking the long-range dependency modeling of transformers. Additionally, parallel strip convolutions are incorporated to efficiently capture anisotropic degradation patterns and directional textures. Furthermore, to resolve the spectral conflict between deep and shallow features within U-Net skip connections, we design the frequency-aware fusion (FaFusion) module. By dynamically decoupling features, this module achieves precise integration of global structure and local texture. Extensive experiments on 10 benchmark datasets across image dehazing, desnowing, and defocus deblurring demonstrate that AKNet consistently outperforms existing methods, achieving a superior balance among reconstruction performance, parameter size, and computational complexity. The source code is available at <a href="https://github.com/PoetryLin/AKNet">https://github.com/PoetryLin/AKNet</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

AKNet: an aligned kernel network for image restoration

  • 期刊论文

  • 发表日期：

    2026年7月23日

查看全文
  • Wan Li,
  • Xiaolin Zhang

摘要

Image restoration aims to recover high-quality visual content from degraded inputs, playing a pivotal role in various real-world applications. While Transformer-based architectures have achieved state-of-the-art performance due to their robust global modeling capabilities, their deployment is severely hindered by inherent quadratic computational complexity. To address this challenge, we propose an efficient aligned kernel network (AKNet). At its core lies the aligned kernel module (AKModule), which innovatively employs super-large convolution kernels to capture global receptive fields with minimal computational overhead, effectively mimicking the long-range dependency modeling of transformers. Additionally, parallel strip convolutions are incorporated to efficiently capture anisotropic degradation patterns and directional textures. Furthermore, to resolve the spectral conflict between deep and shallow features within U-Net skip connections, we design the frequency-aware fusion (FaFusion) module. By dynamically decoupling features, this module achieves precise integration of global structure and local texture. Extensive experiments on 10 benchmark datasets across image dehazing, desnowing, and defocus deblurring demonstrate that AKNet consistently outperforms existing methods, achieving a superior balance among reconstruction performance, parameter size, and computational complexity. The source code is available at https://github.com/PoetryLin/AKNet.