深度可分离残差网络的高分遥感影像场景分类

Depthwise separable residual network for high-resolution remote sensing image scene classification

针对已有基于深度卷积神经网络的高分辨率遥感影像场景分类模型参数量较大和运行速度较慢,以及在训练样本有限的情况下容易出现模型过拟合等问题,该文提出一种基于深度可分离残差网络的高分影像场景分类模型。首先,该模型基于残差学习思想,结合二维卷积和可分离卷积构建残差可分离特征提取模块,能够在提取影像深层次特征的同时减少模型的参数量;然后,以该模块作为基础结构构建深层次的特征提取网络模型;最后,将提取的特征输入到softmax分类器进行分类。在UC Merced和NWPU45两个数据集上对该文方法和几种深度学习场景分类方法进行实验对比,结果表明,提出模型的分类精度在UC Merced数据集两种不同比例训练样本下分别提高到98.04%和99.32%,在NWPU45数据集两种不同比例训练样本下分别提高到90.17%和92.26%,体现出该模型在遥感图像场景分类任务上更具有优势。

In view of the large amount of parameters and slow running speed of the existing high-resolution remote sensing image scene classification models based on deep convolutional neural networks,as well as the tendency to over-fit the model when the training samples are limited,a high-resolution remote sensing image scene classification model based on depthwise separable residual network was proposed in this paper.Firstly,based on the idea of residual learning,the model was combined with two-dimensional convolution and separable convolution to construct a residual separable feature extraction module(RSFE),which could extract the deep features of the image and reduce the number of parameters of the model.Then,RSFE was used as the basic structure to construct a deep feature extraction network model.Finally,the extracted features were input to the softmax classifier for classification.The experimental comparisons between proposed method and other deep learning scene classification methods were carried out on the UC Merced and NWPU45 datasets.The results showed that the classification accuracy of the proposed model was improved to 98.04% and 99.32% under two different proportions of training samples in the UC Merced dataset,and 90.17% and 92.26% under two different proportions of training samples in the NWPU45 dataset,respectively.This model had more advantages in the task of remote sensing image scene classification.