| 引用本文： |
-
张丽丽,王春阳,刘佳辉,等.面向高光谱图像分类的多维协同与中心注意网络[J].电讯技术，2026，66（5）： - . [点击复制]
- ZHANG Lili,WANG Chunyang,LIU Jiahui,et al.Multi-dimensional Synergistic and Central Attention Network for Hyperspectral Image Classification[J].，2026，66（5）： - . [点击复制]
|
|
| 摘要: |
| 现阶段，基于深度学习的分类方法主要依赖于高光谱(HyperSpectral Image，HSI)中的丰富光谱和空间信息。然而，一些方法存在过度使用空间或光谱信息的问题，从而忽视了这两者之间的潜在相关性。为了更有效地结合空间与光谱信息，提出了一种多维协同中心注意(Multi-dimensional Collaboration Center Attention，MCCA)网络。首先，设计了大核多维协同注意模块。该模块能够在通道、高度和宽度3个维度上计算注意力权重，建立多维联系，对关键特征进行提取。其次，采用条件动态位置编码模块通过输入动态调整位置编码，减轻了绝对位置编码在空间信息提取时的干扰。此外，提出了光谱中心注意力模块。该模块能够结合中心像素的光谱信息自适应地提取周围像素的空间相关特征，增强了网络对中心像素的关注能力。该网络在Indian Pines、Pavia University和WHU-Hi-LongKou 3个公共高光谱数据集上的总体准确率分别达到了98.15%、99.09%和99.06%，在训练样本相对较少的情况下取得了良好的分类效果。 |
| 关键词: 高光谱图像分类 多维协同 条件动态位置编码 光谱中心注意力 |
| DOI：10.20079/j.issn.1001-893x.250106001 |
|
| 基金项目:辽宁省教育厅项目（JYTMS20230243） |
|
| Multi-dimensional Synergistic and Central Attention Network for Hyperspectral Image Classification |
| ZHANG Lili,WANG Chunyang,LIU Jiahui,FANG Qizhi |
| (1.College of Electronic Information Engineering，Shenyang Aerospace University，Shenyang 110136，China;2.Liaoning General Aviation Academy，Shenyang 110136，China) |
| Abstract: |
| At present，deep learning-based methods for hyperspectral image classification primarily rely on the rich spectral and spatial information contained in hyperspectral data.However，many existing approaches tend to overutilize either spatial or spectral information，thereby neglecting their potential correlations.To more effectively integrate spatial and spectral features，a multi-dimensional collaborative center attention（MCCA） network is proposed.First，a multi-dimensional collaborative attention module is designed to compute attention weights across channel，height，and width dimensions，enabling multi-dimensional feature extraction and the identification of key features.Second，a dynamic conditional position encoding module is introduced to dynamically adjust position encoding based on input，mitigating the interference caused by absolute position encodings in spatial feature extraction.Additionally，a spectral center attention module is developed to adaptively extract spatially relevant features of neighboring pixels by leveraging the spectral information of the central pixel，thereby enhancing the network’s focus on central pixels.The proposed network achieves overall accuracies of 98.15%，99.09%，and 9906% on the Indian Pines，Pavia University，and WHU-Hi-LongKou hyperspectral datasets，respectively，demonstrating robust classification performance even with relatively limited training samples. |
| Key words: hyperspectral image classification multi-dimensional collaboration dynamic conditional position encoding spectral center attention |
《电讯技术》编辑部 