系统工程与电子技术 ›› 2023, Vol. 45 ›› Issue (9): 2673-2680.doi: 10.12305/j.issn.1001-506X.2023.09.04

• 电子技术 • 上一篇    下一篇

基于三维注意力与混合卷积的高光谱图像分类

赵晓枫1,2, 牛家辉1,2,*, 刘春桐1,2, 夏玉婷1,2   

  1. 1. 火箭军工程大学导弹工程学院, 陕西 西安 710025
    2. 兵器发射理论与技术国家重点学科实验室, 陕西 西安 710025
  • 收稿日期:2022-05-09 出版日期:2023-08-30 发布日期:2023-09-05
  • 通讯作者: 牛家辉
  • 作者简介:赵晓枫 (1979—), 男, 副教授, 博士, 主要研究方向为兵器发射理论与技术
    牛家辉 (1994—), 男, 硕士研究生, 主要研究方向为高光谱图像处理、光电隐身伪装与防护
    刘春桐 (1974—), 男, 教授, 博士, 主要研究方向为定位定向与光电防护
    夏玉婷 (1998—), 女, 硕士研究生, 主要研究方向为图像处理、光电隐身伪装防护
  • 基金资助:
    国家自然科学基金(41404022)

Hyperspectral image classification based on hybrid convolution with three-dimensional attention mechanism

Xiaofeng ZHAO1,2, Jiahui NIU1,2,*, Chuntong LIU1,2, Yuting XIA1,2   

  1. 1. Missile Engineering College, Rocket Force Engineering University, Xi'an 710025, China
    2. Armament Launch Theory and Technology Key Discipline Laboratory of China, Xi'an 710025, China
  • Received:2022-05-09 Online:2023-08-30 Published:2023-09-05
  • Contact: Jiahui NIU

摘要:

针对现有高光谱图像分类模型在特征提取的过程中有效特征关注缺乏的问题, 提出了一种基于三维卷积注意力与混合卷积的高光谱图像分类方法。该方法使用三维卷积和二维卷积串联完成对高光谱图像空谱特征的提取, 并在三维卷积阶段引入注意力机制, 使得模型在提取底层空谱特征的同时实现对有效特征的关注和激活。相对于传统三维卷积模型, 提出的分类模型减小了运算复杂度, 提升了模型噪声抑制能力, 提高了分类效果。针对该方法的消融实验证明了提出的三维卷积注意力机制的有效性, 在印第安松树林和帕维亚大学两个公开数据集上与其他5种分类模型的对比实验中取得了最优的分类精度。

关键词: 高光谱图像分类, 三维注意力机制, 混合卷积, 空谱特征提取

Abstract:

Aiming at the lack of effective attention in the process of feature extraction in existing hyperspectral image classification models, a classification model based on hybrid convolution and three-dimensional attention mechanism is proposed. The method realizes the extraction of spatial-spectral features of hyperspectral images by tandem three-dimensioal (3D) convolution and two-dimensional (2D) convolution. An attention mechanism in the 3D convolution stage is designed, and the attention mechanism is implemented in the 3D convolution stage to realize the attention and activation of the effective spatial-spectral features of hyperspectral images while the model is extracting the underlying features. Compared with the traditional 3D convolution-based model, the classification model proposed in this paper reduces the complexity of operations, improves the model's ability to suppress interference noise, and enhances the classification effect. Ablation experiments against the method demonstrate the effectiveness of the proposed 3D convolution attention mechanism, and the optimal classification accuracy is achieved in comparison experiments with five other classification models on two publicly available datasets, Indian Pines and Pavia University.

Key words: hyperspectral image classification, three-dimensional attention mechanism, hybrid convolution, spatial-spectral feature extraction

中图分类号: