Hyperspectral Image Classification Based on A Multi-Scale Weighted Kernel Network

作     者：SUN Le XU Bin LU Zhenyu SUN Le;XU Bin;LU Zhenyu

作者机构：School of Computer Science Nanjing University of Information Science and Technology Engineering Research Center of Digital Forensics Ministry of Education Nanjing University of Information Science and Technology School of Artificial Intelligence Nanjing University of Information Science and Technology 

出 版 物：《Chinese Journal of Electronics》 (电子学报(英文))

年 卷 期：2022年第31卷第5期

页      面：832-843页

核心收录：

学科分类：0810[工学-信息与通信工程] 08[工学] 081002[工学-信号与信息处理] 

基　　金：supported by the National Natural Science Foundation of China (61971233  U20B2061) 

主　　题：multibranch structures image features convolution kernels multiple convolution kernels adaptive receptive field image classification feature extraction HSI classification single convolution kernel convolutional neural nets hyperspectral image classification representative features hyperspectral imaging geophysical image processing deep learning (artificial intelligence) multiscale weighted kernel network deep learning networks hyperspectral data sets three-branch network HSI cubic patches principal component analysis 1D spectral convolution different-sized convolution kernels 

摘      要：Recently, many deep learning models have shown excellent performance in hyperspectral image(HSI) classification. Among them, networks with multiple convolution kernels of different sizes have been proved to achieve richer receptive fields and extract more representative features than those with a single convolution kernel. However, in most networks, different-sized convolution kernels are usually used directly on multibranch structures, and the image features extracted from them are fused directly and simply. In this paper, to fully and adaptively explore the multiscale information in both spectral and spatial domains of HSI, a novel multi-scale weighted kernel network(MSWKNet) based on an adaptive receptive field is proposed. First, the original HSI cubic patches are transformed to the input features by combining the principal component analysis and one-dimensional spectral convolution. Then, a three-branch network with different convolution kernels is designed to convolve the input features, and adaptively adjust the size of the receptive field through the attention mechanism of each branch. Finally, the features extracted from each branch are fused together for the task of *** on three well-known hyperspectral data sets show that MSWKNet outperforms many deep learning networks in HSI classification.