基于浅层特征调制的轻量级单幅图像超分辨率重建
CSTR:
作者:
作者单位:

中国矿业大学

作者简介:

通讯作者:

中图分类号:

TP391

基金项目:

国家自然科学基金项目(面上项目,重点项目,重大项目)


Lightweight Single Image Super-Resolution Based on Shallow Feature Modulation
Author:
Affiliation:

China University of Mining and Technology

Fund Project:

The National Natural Science Foundation of China (General Program, Key Program, Major Research Plan)

  • 摘要
  • |
  • 图/表
  • |
  • 访问统计
  • |
  • 参考文献
  • |
  • 相似文献
  • |
  • 引证文献
  • |
  • 资源附件
  • |
  • 文章评论
    摘要:

    浅层特征在超分辨率重建网络中扮演关键角色,其中蕴含丰富的图像细节,对准确估计深层特征具有明显参考价值.然而,研究者常常忽视浅层特征,过度依赖深层模块堆叠和拓扑结构优化,进而造成信息冗余.因此,本研究提出一种轻量级超分辨率重建网络,旨在探索浅层特征与深层特征的映射机制,以提升重建质量.首先,通过利用浅层特征生成特征掩码,引导深层特征的生成过程;其次,采用基于注意力机制的特征选择模块,动态生成特征权重信息;最后,设计了双分支特征增强学习模块,平衡输出特征权重并增强特征融合能力,进一步提升重建性能.实验结果表明,所提出的算法在国际通用数据集上显著提升了峰值信噪比和结构相似度指标,同时具有较小的模型参数量和卓越的视觉表现.这些结果验证了本研究所提出的轻量级超分辨率重建网络的有效性和优越性.

    Abstract:

    Shallow features play a pivotal role in super-resolution reconstruction networks, as they encompass intricate image details and offer explicit reference value for precise estimation of deep features. Nonetheless, researchers frequently disregard shallow features and excessively depend on deep module stacking and topology optimization, leading to redundant information. Consequently, this study introduces a lightweight super-resolution reconstruction network that endeavours to explore the mapping mechanism between shallow and deep features to enhance reconstruction quality. Firstly, shallow features are leveraged to generate feature masks that guide the generation process of deep features. Secondly, an attention-based feature selection module dynamically generates feature weight information. Finally, a dual-branch feature enhancement learning module is devised to balance the output feature weights and augment the feature fusion capability, thereby further improving the reconstruction performance. Experimental findings substantiate that the proposed algorithm significantly enhances the peak signal-to-noise ratio and structural similarity metrics on a widely used international dataset, while concurrently exhibiting a limited number of model parameters and exemplary visual performance. These outcomes validate the efficacy and superiority of the lightweight super-resolution reconstruction network proposed in this study.

    参考文献
    相似文献
    引证文献
引用本文
分享
文章指标
  • 点击次数:
  • 下载次数:
  • HTML阅读次数:
  • 引用次数:
历史
  • 收稿日期:2023-12-11
  • 最后修改日期:2024-09-19
  • 录用日期:2024-04-24
  • 在线发布日期: 2024-05-06
  • 出版日期:
文章二维码