Scholarlink

<p>Hyperspectral imaging consists of large quantities of spatial-spectral information. Visual recognition systems driven by deeper and larger convolutional network architectures enable accurate classification of complex scenes. Ensemble Learning methods are popular for enhancing a model’s performance, but such techniques increase the computational requirements several times as model complexity increases. This research proposes a knowledge-distillation framework wherein a more extensive teacher network transfers its spatial-spectral knowledge to a smaller student network. The student network is trained using a combination of distillation and cross-entropy loss. The teacher network learns intricate patterns from the HSI data and produces softened output probabilities that encode interclass relationships. These are distilled into a computationally efficient student network trained with a combined loss function that aligns its predictions with the teacher’s output. Experimental results on three different landcover datasets demonstrate that the distilled student network achieved an accuracy of 96.9% compared to 92.3% with the standalone student network while reducing the model complexity for the agricultural landcover dataset.</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

Distilling spectral-spatial knowledge for efficient hyperspectral image classification

  • 期刊论文

  • 发表日期：

    2025年6月26日

查看全文
  • Vamshi Krishna Munipalle,
  • Usha Rani Nelakuditi,
  • Rama Rao Nidamanuri

摘要

Hyperspectral imaging consists of large quantities of spatial-spectral information. Visual recognition systems driven by deeper and larger convolutional network architectures enable accurate classification of complex scenes. Ensemble Learning methods are popular for enhancing a model’s performance, but such techniques increase the computational requirements several times as model complexity increases. This research proposes a knowledge-distillation framework wherein a more extensive teacher network transfers its spatial-spectral knowledge to a smaller student network. The student network is trained using a combination of distillation and cross-entropy loss. The teacher network learns intricate patterns from the HSI data and produces softened output probabilities that encode interclass relationships. These are distilled into a computationally efficient student network trained with a combined loss function that aligns its predictions with the teacher’s output. Experimental results on three different landcover datasets demonstrate that the distilled student network achieved an accuracy of 96.9% compared to 92.3% with the standalone student network while reducing the model complexity for the agricultural landcover dataset.