Scholarlink

<p>Hyperspectral image (HSI) reconstruction from RGB images offers a cost-effective solution, significantly reducing acquisition costs and expanding application scope. However, existing deep learning-based HSI reconstruction methods struggle with noise, artifacts, and overexposure in RGB images, leading to low reconstruction accuracy and poor local edge preservation. To address these challenges, we propose an Enhanced Hyperspectral Image Reconstruction via Parallel 2D/3D Convolution with Global Layer Purification and Multiscale Pooling Fusion (P2D/3D HRNet). The Global Layer-wise Purification Module (GLPM) effectively eliminates noise in shallow RGB features. The Multiscale Pooling Fusion Module (MPFM) fuses shallow spatial contextual information, local details, and deep spatial-spectral features, enhancing local edge reconstruction. Additionally, an Efficient 3D Convolution Pathway (E3DCP), parallel to the 2D convolution branch, captures sufficient spatial and spectral information, improving spectral accuracy. Experimental results on benchmark datasets demonstrate that our method outperforms most state-of-the-art approaches, achieving a significant reduction in mean relative absolute error 7.82% on the NTIRE2022 dataset and root mean square error 8.15% on the NTIRE2020-Real dataset. This study underscores the potential of our approach for practical hyperspectral imaging applications. The code can be obtained at <a href="https://github.com/zhang-yujie/P2D-3D_HRNet">https://github.com/zhang-yujie/P2D-3D_HRNet</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

Enhanced hyperspectral image reconstruction via parallel 2D/3D convolution with global layer purification and multiscale pooling fusion

  • 期刊论文

  • 发表日期：

    2025年8月18日

查看全文
  • Yujie Zhang,
  • Suyu Wang,
  • Yuqing Cui

摘要

Hyperspectral image (HSI) reconstruction from RGB images offers a cost-effective solution, significantly reducing acquisition costs and expanding application scope. However, existing deep learning-based HSI reconstruction methods struggle with noise, artifacts, and overexposure in RGB images, leading to low reconstruction accuracy and poor local edge preservation. To address these challenges, we propose an Enhanced Hyperspectral Image Reconstruction via Parallel 2D/3D Convolution with Global Layer Purification and Multiscale Pooling Fusion (P2D/3D HRNet). The Global Layer-wise Purification Module (GLPM) effectively eliminates noise in shallow RGB features. The Multiscale Pooling Fusion Module (MPFM) fuses shallow spatial contextual information, local details, and deep spatial-spectral features, enhancing local edge reconstruction. Additionally, an Efficient 3D Convolution Pathway (E3DCP), parallel to the 2D convolution branch, captures sufficient spatial and spectral information, improving spectral accuracy. Experimental results on benchmark datasets demonstrate that our method outperforms most state-of-the-art approaches, achieving a significant reduction in mean relative absolute error 7.82% on the NTIRE2022 dataset and root mean square error 8.15% on the NTIRE2020-Real dataset. This study underscores the potential of our approach for practical hyperspectral imaging applications. The code can be obtained at https://github.com/zhang-yujie/P2D-3D_HRNet.