基于CNN-CBIR的遥感图像分类检索方法  

Remote Sensing Image Classification and Retrieval Method Based on CNN-CBIR

作　　者：马广迪 杨为琛 MA Guangdi;YANG Weichen(Zhejiang Guoyao Geographic Information Technology Company Limited,Huzhou Zhejiang 313200,China)

机构地区：[1]浙江国遥地理信息技术有限公司,浙江湖州313200

出　　处：《北京测绘》2021年第5期634-639,共6页Beijing Surveying and Mapping

摘　　要：遥感图像海量性、复杂性与多样性特征导致现有方法出现查全率、查准率低的问题,无法满足现今遥感图像应用的需求,故提出基于卷积神经网络-图像检索(Convolutional Neural Networks-ContentBased Image Retrieval,CNN-CBIR)的遥感图像分类检索方法研究。为了精确分类遥感图像,基于卷积神经网络-深度卷积神经网络-16 (Convolutional Neural Networks-Visual Geometry Group Net-16,CNN-VGGNet-16)模型提取遥感图像卷积特征与池化特征,通过有效融合得到遥感图像高层聚合特征,以此为基础,采用模糊分类算法分类处理遥感图像,依据遥感图像分类结果,利用基于内容的图像检索(Content-Based Image Retrieval,CBIR)技术制定遥感图像分类检索程序,实现了遥感图像的分类检索。选取数据集图像遥感数据集(UC-Merced)与武大遥感数据集(WHU-RS)作为实验数据集,确定最佳池化区域尺寸与最佳输入尺寸,采用MATLAB软件进行仿真实验。仿真实验数据显示:与标准数值相比较,提出方法的查全率与查准率较高,充分说明提出方法具备更好的检索性能。Due to the characteristics of mass,complexity and diversity of remote sensing images,the existing methods have problems of low recall and precision,which cannot meet the application requirements of current remote sensing images.Therefore,the classification and retrieval method of remote sensing images based on CNN-CBIR was proposed.For accurate classification of remote sensing image based on convolutional neural network(CNN)VGGNet-16 extraction of remote sensing image convolution model features and pooling,remote sensing image was obtained by effective fusion of high-level polymerization characteristics,on this basis,using fuzzy classification processing of remote sensing image classification algorithms,based on remote sensing image classification as a result,the use of CBIR technology for remote sensing image classification retrieval program,realized the classification of remote sensing image retrieval.Uc-merced and WHU-RS data sets were selected as experimental data sets to determine the best pooling area size and the best input size,and MATLAB software was used for simulation experiment.Simulation results showed that compared with the standard values,the proposed method had higher recall and precision,which fully indicated that the proposed method had better retrieval performance.

关 键 词：卷积神经网络(CNN-CBIR) 遥感图像 分类 检索 查全率 

分 类 号：P237[天文地球—摄影测量与遥感]