基于显著性特征和DCNN的高分遥感影像场景分类  

High-Resolution Remote Sensing Scene Classification Based on Salient Features and DCNN

作　　者：吕欢欢[1] 刘涛 张辉[1] 彭国峰 张峻通 Lu Huanhuan;Liu Tao;Zhang Hui;Peng Guofeng;Zhang Juntong(College of Software,Liaoning Technical University,Huludao,Liaoning 125105,China)

机构地区：[1]辽宁工程技术大学软件学院,辽宁葫芦岛125105

出　　处：《激光与光电子学进展》2021年第20期36-48,共13页Laser & Optoelectronics Progress

基　　金：国家自然科学基金(41871379,61540056);辽宁省自然科学基金指导计划(20180550450)。

摘　　要：高分遥感影像的场景分类是解译遥感影像信息的重要工作之一。为了准确提取出目标信息,针对高分遥感影像场景分类中存在的背景复杂、目标多样、目标信息与背景信息难以区分等问题,提出了一种基于显著性特征和深度卷积神经网络(DCNN)的高分遥感影像场景分类方法。首先,利用K-means聚类与超像素分割算法得到影像的颜色空间分布与颜色对比图,融合不同对比图,以得到显著图。然后,通过对数变换增强显著图中的特征,采用自适应阈值分割方法提高目标的区分度并划分出目标区域和背景区域,以提取出感兴趣区域。最后,构建了一种用于提取深层语义特征的DCNN模型,并将得到的特征输入网络模型中进行训练和分类。实验结果表明,本方法能有效区分主要目标信息与背景信息,减少无关信息的干扰,在UC-Merced数据集和WHU-RS数据集上的分类精度分别为96.10%和95.84%。Scene classification of high-resolution remote sensing image is one of the important tasks in interpreting remote sensing image information.In order to extract the target information accurately,we propose a highresolution remote sensing image scene classification method based on salient features combined with deep convolutional neural network(DCNN)to solve the problems of complex background,diverse targets,and difficult to distinguish between target information and background information in the classification of high-scoring remote sensing image scenes.First,K-means clustering algorithm and super-pixel segmentation algorithm are used to generate the color spatial distribution map and color contrast map of the image,and the different contrast maps are fused to get the saliency map.Then,the features in the saliency map are enhanced through logarithmic transformation,and the adaptive threshold segmentation method is used to improve the discrimination of the target and divide the target area and the background area,and extract the area of interest.Finally,a DCNN model is constructed to extract deep semantic features and classification,and the obtained features are input into the network model for training and classification.Experimental results show that the method can effectively distinguish the main target information from the background information and reduce the interference of irrelevant information.The classification accuracy of the method on the UC-Merced data set and WHU-RS data set are 96.10% and 95.84%,respectively.

关 键 词：大气光学 高分遥感影像 场景分类 显著性检测 卷积神经网络 深层语义特征 

分 类 号：P237[天文地球—摄影测量与遥感]