基于改进卷积神经网络的遥感图像目标检测方法  

Method of remote sensing image target detection based on improved convolution neural network

作　　者：王艳辉 张福泉 邹静[3] 侯小毛 Wang Yanhui;Zhang Fuquan;Zou Jing;Hou Xiaomao(School of Computer Science and Engineering,Hunan University of Information Technology,Changsha 410151,China;College of Computer and Control Engineering,Minjiang University,Fuzhou 350108,China;School of Computer Science and Technology,Guizhou University,Guiyang 550025,China)

机构地区：[1]湖南信息学院计算机科学与工程学院,湖南长沙410151 [2]闽江学院计算机与控制工程学院,福建福州350108 [3]贵州大学计算机科学与技术学院,贵州贵阳550025

出　　处：《南京理工大学学报》2023年第3期330-336,共7页Journal of Nanjing University of Science and Technology

基　　金：国家自然科学基金(61871204);福建省科技厅引导性项目(2018H0028)。

摘　　要：为了提高遥感目标检测的稳健性和准确性,基于低层特征检测器,增加了1个改进型卷积神经网络(CNN)框架。首先,利用支持向量回归(SVR)对遥感目标进行初步分类,将检测出的目标信息作为CNN框架的输入。然后,对CNN框架进行优化,通过模块扩展的方式纳入更深的模块。为了使得分类器对亮度变化具有更好的稳健性,在特征向量分类之前增加正则化层(RL)。同时,为了提升目标检测的准确性,增加1个欧拉变换层(ETL),作为类别间的分离度量。使用来自CIFAR-10和MNIST数据集中的图像,与定向梯度边缘直方图(E-HOG)方法、基于生成式对抗网络(GAN)的检测方法、基于二值与浮点数混用方法的语义分割网络(MBU-Net)相比较,仿真结果表明:该文方法的精度和F1得分更高,且标准偏差也更低;该文方法的运行时间接近于一般CNN方法;利用该文方法在测试集的卫星图像中进行目标建筑物检测,模块化CNN可以与基于特征的算法实现互补。To improve the robustness and accuracy of remote sensing target detection,based on low-level feature detector,an improved convolutional neural network(CNN)framework is added.Firstly,support vector regression(SVR)is used to classify the remote sensing target,and the detected target information is used as the input of CNN framework.Then,the CNN framework is optimized and incorporated into deeper modules through module expansion.In order to make the classifier more robust to brightness changes,a regularization network layer(RL)is added before feature vector classification.At the same time,in order to improve the accuracy of target detection,an Euler transform layer(ETL)is added as the separation measure between categories.Using images from CIFAR-10 and MNIST datasets,the method proposed here is compared with the edge-histogram of gradient(E-HOG)method,the GAN based detection method,and the mixed binary U-shape network(MBU-Net)based on the mixed binary and floating point number method.The simulation results show that the accuracy and F1 scores of this method are higher,and the standard deviation is lower;the running time of this method is close to that of general CNN method;using this method to detect target buildings in satellite images of test sets,modular CNN can complement feature-based algorithms.

关 键 词：卷积神经网络 遥感图像 目标检测 支持向量回归 欧拉变换层 卫星图像 建筑物检测 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]