Scholarlink

<p>In recent years, the advantage of convolutional neural networks in capturing local spatial information has achieved significant performance in single-image super-resolution. However, most methods typically use small convolution kernels to aggregate local and global features, thereby ignoring the interactions between channels and spatial dimensions within larger receptive fields. To address this issue, we propose a hybrid feature enhancement network (HFEN) for image super-resolution reconstruction. Specifically, we have designed a large-kernel-transposed attention module that applies self-attention to the channel dimension while expanding the receptive field, effectively capturing channel information and long-range dependencies. In addition, to utilize the rich feature representations in feedforward networks, we designed a multi-scale aggregation feedforward network to extract multi-scale information and aggregate features. Experiments have shown that HFEN is significantly superior to state-of-the-art lightweight SR methods. Code is available at <a href="https://github.com/SJHunag/HFEN">https://github.com/SJHunag/HFEN</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

Hybrid feature enhancement network for lightweight image super-resolution

  • 期刊论文

  • 发表日期：

    2025年4月24日

查看全文
  • Shigang Hu,
  • Shijun Huang,
  • Jianxin Wang

摘要

In recent years, the advantage of convolutional neural networks in capturing local spatial information has achieved significant performance in single-image super-resolution. However, most methods typically use small convolution kernels to aggregate local and global features, thereby ignoring the interactions between channels and spatial dimensions within larger receptive fields. To address this issue, we propose a hybrid feature enhancement network (HFEN) for image super-resolution reconstruction. Specifically, we have designed a large-kernel-transposed attention module that applies self-attention to the channel dimension while expanding the receptive field, effectively capturing channel information and long-range dependencies. In addition, to utilize the rich feature representations in feedforward networks, we designed a multi-scale aggregation feedforward network to extract multi-scale information and aggregate features. Experiments have shown that HFEN is significantly superior to state-of-the-art lightweight SR methods. Code is available at https://github.com/SJHunag/HFEN.