首页|结合组像素嵌入的双注意力高光谱图像分类

结合组像素嵌入的双注意力高光谱图像分类

扫码查看
原文链接
  • 国家科技期刊平台
  • NETL
  • NSTL
  • 万方数据
近年来，基于深度学习的框架在高光谱图像分类领域中取得了令人满意的结果。然而，多数方法仍使用卷积神经网络作为主干网络，其存在感受野过小，对特征信息的挖掘不充分，序列建模的能力较弱，模型复杂和分类精度低等问题。为克服上述局限性，该文提出一种结合组像素嵌入的双注意力高光谱图像分类的方法。该方法主要分成三个部分，首先，使用含有点卷积组和深度卷积组的通道空间卷积分离模块来高效学习空间光谱的特征信息;其次，添加通道空间双注意力机制，抑制冗余信息的干扰，增强高光谱图像空间与光谱的特征权重;最后，通过组像素嵌入Transformer来进一步强化空间与光谱之间的联系，建立全局长距离依赖关系，缓解精度下降的问题，保证了网络良好的分类性能。实验结果表明，该方法与现有的网络模型相比具有更优越的性能，在Pavia University和WHU-Hi-LongKou两个数据集中的总体准确率分别达到99。26%和99。73%。
Hyperspectral Image Classification Using Dual Attention with Grouped Pixel Embedding
In recent years,frameworks based on deep learning have yielded satisfactory results in hyperspectral image classification.However,most methods still rely on convolutional neural networks as the backbone network,which suffer from limitations such as small receptive fields,insufficient exploration of feature information,weak sequential modeling capabilities,complex models,and low classification accuracy.To address these limitations,we propose a method that hyperspectral image classification using dual attention with grouped pixel embedding.The proposed method is divided into three main parts.Firstly,it achieves efficient learning of spatial-spectral feature information through channel spatial separation convolution module comprising point convolution groups and depth convolution groups.Secondly,it incorporates a channel spatial dual attention mechanism to suppress interference from redundant information and enhance the feature weights of spatial and spectral characteristics in hyperspectral images.Finally,it utilizes a grouped pixel embedding transformer to strengthen the relationship between spatial and spectral information,establish global long-distance dependency relationships,alleviate accuracy degradation issues,and ensure the network's robust classification performance.Experimental results dem-onstrate that the proposed method outperforms existing network models,achieving overall accuracies of 99.26%and 99.73%on the Pavia University and WHU-Hi-LongKou datasets,respectively.

hyperspectral image classificationconvolutional neural networkschannel spatial separation convolutiondual attention mechanismgrouped pixel embedding Transformer

谭云飞、李明、罗勇航、石超山、文贵豪

展开 >

重庆师范大学 计算机与信息科学学院,重庆 401331

高光谱图像分类 卷积神经网络 通道空间卷积分离 双注意力机制 组像素嵌入Transformer

国家自然科学基金国家自然科学基金重庆市教委项目重庆市研究生教改重点项目

6187705161170192113143yjg182022

2024

10.20165/j.cnki.ISSN1673-629X.2024.0158

计算机技术与发展
陕西省计算机学会

计算机技术与发展

CSTPCD
影响因子：0.621
ISSN：1673-629X
年,卷(期)：2024.34(9)