多尺度特征与双注意力机制的高光谱影像分类

Multi-scale feature and dual-attention mechanism for hyperspectral image classification

针对经典卷积神经网络(convolutional neural network, CNN)的高光谱影像分类方法存在关键细节特征表现不足、训练需要大量样本等问题,提出一种基于多尺度特征与双注意力机制的高光谱影像分类方法。首先,利用三维卷积提取影像的空谱特征,并采用转置卷积获得特征的细节信息;然后,通过不同尺寸的卷积核运算提取多尺度特征并实现不同感受野下多尺度特征的融合;最后,设计双注意力机制抑制混淆的区域特征,同时突出区分性特征。在两幅高光谱影像上进行的实验结果表明:分别在每类地物中随机选取10%和0.5%的样本作为训练样本,提出模型的总体分类精度分别提高到99.44%和98.86%;对比一些主流深度学习分类模型,提出模型能够关注于对分类任务贡献最大的关键特征,可以获取更高的分类精度。

Hyperspectral image classification methods based on the classical convolutional neural network(CNN)have some problems,such as insufficient expression of key detail features and a large number of samples for training.Aiming at these problems,this paper proposes a hyperspectral image classification model with multi-scale features and dual-attention mechanism.Firstly,using 3D convolution,the spatial-spectral features of images can be directly extracted,and transposed convolution is adopted to get more details of the feature map.Then,a feature extraction module is built through convolution kernels of different sizes to achieve multi-scale feature fusion under different receptive fields.Finally,the dual-attention mechanism is designed to suppress the confused regional features and highlight the distinguishing features.The experimental results on two hyperspectral images show that when 10%and 0.5%samples are randomly selected as training samples for each class of ground object,the overall classification accuracy of the proposed model is improved to 99.44%and 98.86%,respectively.This model can obtain higher classification accuracy than some mainstream deep-learning classification models.Since the model can focus on more important detailed features during feature extraction,the classification effect is improved.