Scholarlink

<p>Multimodal image registration and fusion are crucial in computer vision, particularly for infrared and visible image processing. However, existing deep learning-based methods face significant challenges: registration approaches struggle with cross-modal feature matching as they are confined to a single-image domain, while fusion techniques often suffer from information loss. To address these issues, we propose a Dual-Pyramid Framework for Robust Infrared-Visible Image Registration and Fusion (DPRF). Our approach utilizes a Gaussian-Dictionary Dual Encoding Residual Deformation Field Registration Network (GDFR-Net) for accurate cross-modal registration. GDFR-Net jointly optimizes feature learning in both the image and dictionary domains, mitigating modality differences. A layer-wise residual deformation field estimation strategy further refines registration accuracy. For fusion, we introduce the Laplacian Pyramid and Dictionary Learning-based Multi-scale Feature Fusion Network (LPDF-Net). LPDF-Net integrates Laplacian pyramid decomposition with convolutional dictionary learning to adaptively reconstruct low-frequency components, preserving infrared saliency and visible contrast. Meanwhile, a dynamic weighting mechanism enhances high-frequency details, ensuring sharp textures and structural integrity. Extensive experiments demonstrate the state-of-the-art performance of our DPRF, effectively reducing misalignment artifacts and producing visually natural fused images, underscoring its potential for real-world applications. The source code for the DPRF is publicly available at <a href="https://github.com/fdaijdfina/DPRF">https://github.com/fdaijdfina/DPRF</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

Dual-pyramid framework for robust infrared-visible image registration and fusion via convolutional dictionary learning

  • 期刊论文

  • 发表日期：

    2025年9月15日

查看全文
  • Chaojie Wu,
  • Zheng Li

摘要

Multimodal image registration and fusion are crucial in computer vision, particularly for infrared and visible image processing. However, existing deep learning-based methods face significant challenges: registration approaches struggle with cross-modal feature matching as they are confined to a single-image domain, while fusion techniques often suffer from information loss. To address these issues, we propose a Dual-Pyramid Framework for Robust Infrared-Visible Image Registration and Fusion (DPRF). Our approach utilizes a Gaussian-Dictionary Dual Encoding Residual Deformation Field Registration Network (GDFR-Net) for accurate cross-modal registration. GDFR-Net jointly optimizes feature learning in both the image and dictionary domains, mitigating modality differences. A layer-wise residual deformation field estimation strategy further refines registration accuracy. For fusion, we introduce the Laplacian Pyramid and Dictionary Learning-based Multi-scale Feature Fusion Network (LPDF-Net). LPDF-Net integrates Laplacian pyramid decomposition with convolutional dictionary learning to adaptively reconstruct low-frequency components, preserving infrared saliency and visible contrast. Meanwhile, a dynamic weighting mechanism enhances high-frequency details, ensuring sharp textures and structural integrity. Extensive experiments demonstrate the state-of-the-art performance of our DPRF, effectively reducing misalignment artifacts and producing visually natural fused images, underscoring its potential for real-world applications. The source code for the DPRF is publicly available at https://github.com/fdaijdfina/DPRF.