基于生成对抗网络的半监督遥感图像场景分类  

SEMI-SUPERVISED REMOTE SENSING IMAGE SCENE CLASSIFICATION BASED ON GENERATIVE ADVERSARIAL NETWORKS

作　　者：郭东恩[1] 吴泽琛 GUO Dong-en;WU Ze-chen(School of Computer and Software,Nanyang Institute of Technology,Nanyang 473004,China)

机构地区：[1]南阳理工学院计算机与软件学院,河南南阳473004

出　　处：《南阳理工学院学报》2022年第6期53-59,共7页Journal of Nanyang Institute of Technology

基　　金：河南省科技攻关项目(212102210492);南阳市科技攻关项目(KJGG102)。

摘　　要：提出使用生成对抗网络(generative adversarial networks, GAN)的半监督遥感图像场景分类模型,该模型引入了密集残差块、预训练的ResNet50 V1网络及特征融合思想、金字塔卷积和谱归一化增强判别网络的特征表示能力,以提高半监督分类性能。密集残差块的引入可以提高生成网络生成图像的质量,同时增加判别网络提取特征的判别能力;通过微调预训练的ResNet50 V1网络提取输入图像的语义特征,并与原始判别网络的特征融合进一步增强特征判别能力;金字塔卷积利用不同类型、大小和深度的滤波器来捕获图像中不同层次的细节;谱归一化能够稳定GAN的训练过程从而提高分类精度。在公开的EuroSAT和UC Merced数据集上的实验结果表明,提出的模型获得了最高的分类精度,尤其在仅有少量有标签样本的情况下取得了明显更高的精度。Compared with the field of natural images, the insufficient number of tagged remote sensing images becomes the bottleneck of supervised scene classification methods, while unsupervised methods are difficult to meet the practical needs.To this end, a generative adversarial networks(GAN)-based semi-supervised remote sensing image scene classification model is proposed, which introduces dense residual block, pretrained ResNet50 v1 networks and feature fusion ideas, pyramidal convolution, and spectral normalization into the discriminative networks to enhance the feature representation capability to improve semi-supervised classification performance.Specifically, the semantic features of the input image are extracted by fine-tuning the pretrained ResNet50 V1 network and fused with the features of the original discriminative network to improve the feature discrimination ability;the pyramidal convolution is used to capture different levels of details in the scene image using different types, sizes and depths of filters;and spectral normalization is used to stabilize the training process of GAN to improve classification accuracy.Extensive experimental results on EuroSAT and UC Merced datasets show that the proposed model achieves the highest classification accuracy, especially when only a few tagged samples are available.

关 键 词：遥感图像场景分类 半监督学习 生成对抗网络 金字塔卷积 谱归一化 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]