| 唐婷,潘新.一种基于可变形卷积的高光谱图像分类算法[J].光电子激光,2022,33(5):488~494 |
| 一种基于可变形卷积的高光谱图像分类算法 |
| A hyperspectral image classification algorithm based on deformable convolution |
| 投稿时间：2021-08-13 |
| DOI： |
| 中文关键词: 高光谱图像分类 可变形卷积 深度学习 卷积神经网络 |
| 英文关键词:hyperspectral image classification deformable convolution deep learni ng convolutional neural network |
| 基金项目:国家自然基金(61962048,61562067)和中央级基本科研业务费(1610332020020)资助项目 (内蒙古农业大学计算机与信息工程学院,内蒙古呼和浩特 010018) |
|
| 摘要点击次数: 264 |
| 全文下载次数: 10 |
| 中文摘要: |
| 随着深度学习的不断发展,基于深度学习的机器 视觉方法被广泛应用,其中,卷积神 经网络(convolutional neural network,CNN)对高光谱图像(hyperspectral imagery,HSI ) 分类有着显著的效果。传统卷积网络中卷积核的采样位置是固定的,不能根据HSI中复杂的 空间结构而改变,忽略了数据在空间分布上的特征,为了提高高光谱图像分类在实际应用中 的性能,本文提出了一种基于可变形卷积的高光谱图像分类方法,考虑到HSI高维度的特性 , 将可变形卷积从2D引伸到3D,从而更好地提取3D空间上的特征。本文结合双分支双注意机制网络(double- branch dual-attention mechanism network,DBDA)的网络结构和3D可变形卷积,在Indi an Pines(IP)和Botswana(BS)2个数据集上进行了实验。实验结果表明,本文的方法在 综合精度(overall accuracy, OA) 、平均精度(average accuracy, AA)、KAPPA评价标准上均获得了更好的分类准确 率,相较于次优算法,OA提高了0.15%—0.23%,AA提高了0.21%, KAPPA提高了0.000—0.001。 |
| 英文摘要: |
| With the continuous development of deep learning,machine vision methods based on deep learning are widely used,and the convolutional neural network (CNN ) has remarkable effect on hyperspectral imagery (HSI) classification.The sampling po sition of the convolutional kernel in traditional convolutional networks is fixed and cann ot be changed according to the complex spatial structure in HSI,ignoring the features of the data on spatial distribution.To improve the performance of hyperspectral image classification i n practical applications,this paper proposes a deformable convolution-based hyperspectral image classification method,which extends the deformable convolution from 2D to 3D co nsidering the high dimensionality of HSI,so as to better extract the features on 3D space .This paper combines the double-branch dual-attention mechanism network (DBDA) structure and 3D deformable convolution,and experiments are conducted on two da tasets, Indian Pines (IP) and Botswana (BS),and the experimental results show that the method of this paper achieves better classification accuracy on overall accu- racy (OA),average accuracy (AA) and KAPPA evaluati on criteria, and improves OA by 0.15%—0.23%,AA by 0.21%,and KAPPA by 0.0003—0.0014compared with the suboptimal algorithm. |
| 查看全文 下载PDF阅读器 |
| 关闭 |
|
|
|
|
|