Scholarlink

<p>Image motion blur recovery aims to restore images degraded by camera shake or object motion. Traditional methods often struggle with complex degradation. While convolutional neural networks (CNNs) and Transformers show high performance, they do not meet the needs of resource-constrained environments due to excessive computational overhead or limited receptive fields. To address these challenges, we propose DCAFNet, a lightweight network with an asymmetric encoder-decoder architecture. The network integrates the Dynamic Context-Aware Fusion Module (DCAFM) and the Frequency-Adaptive Attention Block (FAAB). DCAFM enables adaptive extraction and fusion of multi-scale features by employing large convolutional kernels and a gating mechanism, which enhances the model’s representational capacity. FAAB jointly learns features in both frequency and spatial domains, thereby improving high-frequency detail recovery and enhancing overall restoration quality. Experimental results demonstrate that DCAFNet, with just 6.05 MB of parameters, achieves state-of-the-art performance on several public benchmarks, demonstrating its efficiency and practical applicability. The source code is available at <a href="https://github.com/Reg95/DCAFNet">https://github.com/Reg95/DCAFNet</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

DCAFNet: a lightweight network with dynamic context-aware fusion for image deblurring

  • 期刊论文

  • 发表日期：

    2026年4月17日

查看全文
  • Gang Ke,
  • Sio-Long Lo,
  • Hua Zou

摘要

Image motion blur recovery aims to restore images degraded by camera shake or object motion. Traditional methods often struggle with complex degradation. While convolutional neural networks (CNNs) and Transformers show high performance, they do not meet the needs of resource-constrained environments due to excessive computational overhead or limited receptive fields. To address these challenges, we propose DCAFNet, a lightweight network with an asymmetric encoder-decoder architecture. The network integrates the Dynamic Context-Aware Fusion Module (DCAFM) and the Frequency-Adaptive Attention Block (FAAB). DCAFM enables adaptive extraction and fusion of multi-scale features by employing large convolutional kernels and a gating mechanism, which enhances the model’s representational capacity. FAAB jointly learns features in both frequency and spatial domains, thereby improving high-frequency detail recovery and enhancing overall restoration quality. Experimental results demonstrate that DCAFNet, with just 6.05 MB of parameters, achieves state-of-the-art performance on several public benchmarks, demonstrating its efficiency and practical applicability. The source code is available at https://github.com/Reg95/DCAFNet.