Scholarlink

<p>Hyperspectral images (HSIs) have abundant frequency-domain information, but deep learning methods usually focus on spatial domain or spectral domain feature extraction, and rarely consider the explicit extraction of frequency-domain information. This paper proposes a multi-feature fusion network based on wavelet transform and cross-layer guided attention mechanism (WTCGA) for HSI classification. Specifically, a network combining wavelet transform and cross-layer guided attention mechanism is proposed, in which the wavelet transform completes the extraction of frequency-domain features of HSIs and generates multi-scale and multi-directional features. These features can be used by the cross-layer guided attention mechanism to further capture and enhance these spectral and spatial information at different scales, and transmit the attention information from shallow to deep layers to better capture the rich multi-scale fusion features. At the same time, we incorporate a reverse diffusion process after dimensionality reduction to optimize the shallow latent features before feature extraction. Experimental results on three datasets show that compared with other advanced methods, the classification results of the proposed network are highly competitive.</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

Hyperspectral image classification based on a multi-feature fusion network

  • 期刊论文

  • 发表日期：

    2026年6月25日

查看全文
  • Yi Liu,
  • Yujie Yan,
  • Caihong Mu,
  • Xinyu He,
  • Binyang Ma

摘要

Hyperspectral images (HSIs) have abundant frequency-domain information, but deep learning methods usually focus on spatial domain or spectral domain feature extraction, and rarely consider the explicit extraction of frequency-domain information. This paper proposes a multi-feature fusion network based on wavelet transform and cross-layer guided attention mechanism (WTCGA) for HSI classification. Specifically, a network combining wavelet transform and cross-layer guided attention mechanism is proposed, in which the wavelet transform completes the extraction of frequency-domain features of HSIs and generates multi-scale and multi-directional features. These features can be used by the cross-layer guided attention mechanism to further capture and enhance these spectral and spatial information at different scales, and transmit the attention information from shallow to deep layers to better capture the rich multi-scale fusion features. At the same time, we incorporate a reverse diffusion process after dimensionality reduction to optimize the shallow latent features before feature extraction. Experimental results on three datasets show that compared with other advanced methods, the classification results of the proposed network are highly competitive.