一种基于交叉注意力网络的高光谱影像分类方法

A cross-attention network for hyperspectral image classification

针对现有高光谱影像分类方法存在全局和局部特征提取不充分、特征融合效率低等问题,该文提出一种能有效提取全局特征和局部特征的双支交叉注意力网络(DCNet)。先通过超像素采样网络(SSN)初步提取全局特征,并输入到Transformer编码器进行特征增强;再通过两次卷积核尺寸不同的三维卷积模块提取局部特征;最后,以交叉注意力机制融合全局特征与局部特征,基于融合后的特征获得分类结果。为防止网络出现梯度消失的情况,DCNet中设计了大量的残差连接模块。为验证该方法的分类效果,选择Pavia University数据集和WHU-Hi-HongHu数据集作为测试数据,将其与另外6种模型的结果进行定量评价和目视判读。实验表明DCNet在两组数据集的总体分类精度分别达到99.44%和98.65%,明显高于另外6种模型。

This paper proposes a dual cross-attention network(DCNet)that can effectively extract global and local features,in response to the problems of insufficient global and local feature extraction and low feature fusion efficiency in existing hyperspectral image classification methods.Firstly,preliminary global features are extracted through superpixel sampling network(SSN)and input into Transformer encoder for feature enhancement;Extract local features through two 3D convolution modules with different kernel sizes;Finally,the cross attention mechanism is used to fuse global and local features,and the classification results are obtained based on the fused features.To prevent gradient vanishing in the network,DCNet has designed a large number of residual connection modules.To verify the classification performance of this method,the Pavia University dataset and WHU-Hi-HongHu dataset were selected as test data,and their results were quantitatively evaluated and visually interpreted with the results of six other models.The experiment shows that the overall classification accuracy of DCNet on two datasets reaches 99.44%and 98.65%,respectively,which is significantly higher than the other six models.