Scholarlink

<p>Most existing infrared and visible image fusion methods still rely mainly on deep spatial-domain representations. Their characterization of high-frequency details, low-frequency structures, and cross-modal frequency distribution consistency remains insufficient, which may lead to detail loss and unstable structural reconstruction in fused results. To address these issues, this paper proposes an infrared and visible image fusion method with joint frequency decomposition and deep feature learning, namely JFDFusion. First, JFDFusion constructs a Non-downsampling Enhanced Discrete Wavelet Transform (N-DEWT) module, which performs robust multi-scale frequency-domain modeling through learnable sub-band mixing and an adaptive soft-thresholding strategy while avoiding information loss caused by downsampling. Then, a lightweight Gating Convolutional Mamba (GCMam) module is designed by combining convolutional gating with a multi-directional selective-scan state space model, so as to capture long-range dependencies and strengthen cross-modal structural alignment. In addition, a Dynamic Weight Adjustment Fusion (DWAF) module is introduced to automatically balance infrared saliency and visible-detail contributions through spatially adaptive weights. Finally, a decoder composed of <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(3\times 3\)</EquationSource> </InlineEquation> and <InlineEquation ID="IEq2"> <EquationSource Format="TEX">\(1\times 1\)</EquationSource> </InlineEquation> convolutional layers reconstructs the fused image. Experiments on three commonly used datasets, including MSRS, Road-Scene, and TNO, demonstrate that JFDFusion achieves competitive subjective visual quality and objective evaluation performance. The proposed method can stably preserve structural information and highlight key infrared targets across different scenes, providing high-quality inputs for subsequent visual perception tasks. The source code is publicly available at <a href="https://github.com/laoyu072466-cyber/JFDFusion">https://github.com/laoyu072466-cyber/JFDFusion</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

JFDFusion: joint frequency decomposition and deep feature learning for infrared–visible image fusion

  • 期刊论文

  • 发表日期：

    2026年7月20日

查看全文
  • Qiucheng Yu,
  • Dongmei Zhou,
  • Lixuan Xiao,
  • Bingmeng Zhu,
  • Shengbing Chen

摘要

Most existing infrared and visible image fusion methods still rely mainly on deep spatial-domain representations. Their characterization of high-frequency details, low-frequency structures, and cross-modal frequency distribution consistency remains insufficient, which may lead to detail loss and unstable structural reconstruction in fused results. To address these issues, this paper proposes an infrared and visible image fusion method with joint frequency decomposition and deep feature learning, namely JFDFusion. First, JFDFusion constructs a Non-downsampling Enhanced Discrete Wavelet Transform (N-DEWT) module, which performs robust multi-scale frequency-domain modeling through learnable sub-band mixing and an adaptive soft-thresholding strategy while avoiding information loss caused by downsampling. Then, a lightweight Gating Convolutional Mamba (GCMam) module is designed by combining convolutional gating with a multi-directional selective-scan state space model, so as to capture long-range dependencies and strengthen cross-modal structural alignment. In addition, a Dynamic Weight Adjustment Fusion (DWAF) module is introduced to automatically balance infrared saliency and visible-detail contributions through spatially adaptive weights. Finally, a decoder composed of \(3\times 3\) and \(1\times 1\) convolutional layers reconstructs the fused image. Experiments on three commonly used datasets, including MSRS, Road-Scene, and TNO, demonstrate that JFDFusion achieves competitive subjective visual quality and objective evaluation performance. The proposed method can stably preserve structural information and highlight key infrared targets across different scenes, providing high-quality inputs for subsequent visual perception tasks. The source code is publicly available at https://github.com/laoyu072466-cyber/JFDFusion.