基于混合卷积与三重注意力的高光谱图像分类网络

针对高光谱图像光谱维度高、现有网络无法提供深度级的多层次特征，从而影响分类精度和速度的问题。首先采用核主成分分析对高光谱图像进行降维，使降维后的数据具有最佳区分度，提出了一种基于混合卷积与三重注意力的卷积神经网络(hybrid convolutional neural network with triplet attention, HCTA-Net)模型，该模型设计了一种基于三维、二维和一维卷积的混合卷积神经网络，通过不同维度卷积神经网络的融合，提取高光谱图像精细的光谱–空间联合特征。在二维卷积中加入深度可分离卷积，减少了模型参数，同时引入三重注意力机制，使用三分支结构实现跨维度信息交互，抑制无用的特征信息。在Indian Pines、Salinas和Pavia University数据集上的实验结果表明，本文提出的模型优于其他对比方法，总体分类精度分别达到了99.16%、99.87%和99.76%。

Hyperspectral image classification based on hybrid convolutional neural network with triplet attention

To solve the problems of a high spectral dimension of hyperspectral images and the failure of the existing network to provide multilevel features at the depth level, which affects the classification accuracy and speed, the kernel principal component analysis is used to reduce the dimensionality of hyperspectral images to have the best data differentiation after dimensionality reduction, and a hybrid convolutional neural network with triplet attention (HCTA-Net) model is proposed to design a hybrid model based on 3D, 2D, and 1D CNN to extract the fine spectral–spatial joint features through the fusion of different dimension convolutions. The model also adds depthwise separable convolution into the 2D-CNN to reduce the model parameters and simultaneously introduces a triplet attention mechanism, which uses a three-branch structure to achieve cross-dimensional information interaction to inhibit useless feature information. Experimental results on the Indian Pines, Salinas, and Pavia University datasets show that the proposed model is superior to other comparison methods, and the overall classification accuracy reaches 99.16%, 99.87, and 99.76%, respectively.