结合注意力机制的轻量化网络高光谱图像分类

Hyperspectral image classification based on lightweight network with attention mechanism

关键词:  高光谱图像分类  多尺度提取模块  注意力机制  轻量化网络  非对称卷积  深度可分离卷积

Key words:  hyperspectral image classification  multi-scale extraction module  attention mechanism  lightweight network  asymmetric convolution  depthwise separable convolution

摘要:

高光谱图像分类是遥感领域的核心难题,其挑战主要源于高光谱图像的光谱维度高、空间分辨率低,导致分类任务中难以充分提取高光谱图像的空谱特征。针对现有的卷积神经网络（CNN）高光谱图像分类模型参数量大、计算资源消耗高、分类精度不足的问题,提出了一种基于注意力机制的轻量化网络高光谱图像分类模型（AMLW-CNN）。为了增强网络的特征提取能力,将空谱特征提取模块设计为2个多尺度提取模块,并将空间特征提取模块的各卷积层通过残差结构连接,同时引入注意力机制来强化网络对有效特征的提取。另外,为了减少模型参数量,采用非对称卷积来替代三维卷积核、深度可分离卷积来替换二维卷积核。实验结果表明:AMLW-CNN的分类精度优于对比算法,计算复杂度更低,鲁棒性更强。在Indian Pines、Salinas和Pavia U 3个数据集上的总体分类精度分别达到了98.5%、99.8%、99.9%。

Abstract:

Hyperspectral image (HSI) classification is a challenging task in the field of remote sensing, because the HSI has high spectral dimensionality and low spatial resolution, which makes it difficult to fully extract the spatial-spectral features of hyperspectral images in the classification task. Aiming at solving the problems of the existing convolutional neural network (CNN)-based HSI classification models, such as large parameter size, high computational cost and low classification accuracy, a lightweight network hyperspectral image classification model based on attention mechanism (AMLW-CNN) is proposed in this paper. In order to enhance feature extraction ability of the network, the spatial-spectral feature extraction module is designed based on two multiscale extraction modules. In addition, we use the residual structures to connect the convolutional layers of spatial feature extraction module and incorporate the attention mechanism to enhance the extraction of useful features. Furthermore, to reduce the number of model parameters, an asymmetric convolution and a depthwise separable convolution are introduced to replace the 3D and 2D convolution kernels, respectively. The experimental results show that classification accuracy of AMLW-CNN is better than that of the comparison algorithms, with lower computational complexity and higher robustness. The overall classification accuracies on the datasets of Indian Pines, Salinas and Pavia U has attain 98.5%,99.8% and 99.9%, respectively.