Scholarlink

<p>As intelligent perception advances from object recognition toward attribute understanding, recognizing surface material-related attributes in vehicle-mounted hyperspectral scenes has become an important problem in hyperspectral perception. Real-world vehicle-mounted hyperspectral scenes are affected by reflections, occlusions, surface contamination, spectral variation, and boundary mixing, which hinder stable material-related recognition. This study proposes a joint framework that integrates continuous wavelength modeling, material-prototype-guided spectral–contextual disentanglement, and mixture-region-oriented soft decision-making. The framework was evaluated against SSRN, HybridSN, RU-Net, DeepLabV3+, LO-SST, and HyperFree on the public HSI-Drive and HyKo2-VIS datasets, which are vehicle-mounted scene-understanding datasets with coarse annotations containing limited vehicle- and surface-related material cues, under a unified protocol with five repeated runs. The evaluation covered conventional segmentation performance, boundary quality, prediction calibration, spectral perturbation consistency, material-feature separability, computational complexity, and stability under controlled Gaussian-noise and missing-band perturbations. On HSI-Drive, the proposed method achieves an ECE of 0.052 and a PSI of 1.77. On HyKo2-VIS, it achieves an ECE of 0.063, an SPC of 0.829, and a PSI of 1.66. Although the proposed method does not consistently outperform the strongest segmentation-oriented baseline in mIoU and Boundary F1, it achieves the lowest ECE and the highest PSI on both datasets. Under spectral-band missing conditions, it also maintains competitive or slightly higher SPC at moderate and high missing rates. These results indicate that the framework mainly improves prediction calibration, material-feature separability, and spectral-domain stability, while its benefits for overall segmentation accuracy remain limited and dataset-dependent.</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

An adaptive deep learning framework for material-related attribute recognition in vehicle-mounted hyperspectral scenes

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月31日

查看全文
  • Siyuan Shen,
  • Wei Su

摘要

As intelligent perception advances from object recognition toward attribute understanding, recognizing surface material-related attributes in vehicle-mounted hyperspectral scenes has become an important problem in hyperspectral perception. Real-world vehicle-mounted hyperspectral scenes are affected by reflections, occlusions, surface contamination, spectral variation, and boundary mixing, which hinder stable material-related recognition. This study proposes a joint framework that integrates continuous wavelength modeling, material-prototype-guided spectral–contextual disentanglement, and mixture-region-oriented soft decision-making. The framework was evaluated against SSRN, HybridSN, RU-Net, DeepLabV3+, LO-SST, and HyperFree on the public HSI-Drive and HyKo2-VIS datasets, which are vehicle-mounted scene-understanding datasets with coarse annotations containing limited vehicle- and surface-related material cues, under a unified protocol with five repeated runs. The evaluation covered conventional segmentation performance, boundary quality, prediction calibration, spectral perturbation consistency, material-feature separability, computational complexity, and stability under controlled Gaussian-noise and missing-band perturbations. On HSI-Drive, the proposed method achieves an ECE of 0.052 and a PSI of 1.77. On HyKo2-VIS, it achieves an ECE of 0.063, an SPC of 0.829, and a PSI of 1.66. Although the proposed method does not consistently outperform the strongest segmentation-oriented baseline in mIoU and Boundary F1, it achieves the lowest ECE and the highest PSI on both datasets. Under spectral-band missing conditions, it also maintains competitive or slightly higher SPC at moderate and high missing rates. These results indicate that the framework mainly improves prediction calibration, material-feature separability, and spectral-domain stability, while its benefits for overall segmentation accuracy remain limited and dataset-dependent.