遥感图像飞机目标分类的卷积神经网络方法

发布时间：2018-01-22 06:28

本文关键词： 可见光遥感飞机分类深度学习卷积神经网络　出处：《中国图象图形学报》2017年05期 　论文类型：期刊论文

【摘要】：目的遥感图像飞机目标分类,利用可见光遥感图像对飞机类型进行有效区分,对提供军事作战信息有重要意义。针对该问题,目前存在一些传统机器学习方法,但这些方法需人工提取特征,且难以适应真实遥感图像的复杂背景。近年来,深度卷积神经网络方法兴起,网络能自动学习图像特征且泛化能力强,在计算机视觉各领域应用广泛。但深度卷积神经网络在遥感图像飞机分类问题上应用少见。本文旨在将深度卷积神经网络应用于遥感图像飞机目标分类问题。方法在缺乏公开数据集的情况下,收集了真实可见光遥感图像中的8种飞机数据,按大致4∶1的比例分为训练集和测试集,并对训练集进行合理扩充。然后针对遥感图像与飞机分类的特殊性,结合深度学习卷积神经网络相关理论,有的放矢地设计了一个5层卷积神经网络。结果首先,在逐步扩充的训练集上分别训练该卷积神经网络,并分别用同一测试集进行测试,实验表明训练集扩充有利于网络训练,测试准确率从72.4%提升至97.2%。在扩充后训练集上,分别对经典传统机器学习方法、经典卷积神经网络Le Net-5和本文设计的卷积神经网络进行训练,并在同一测试集上测试,实验表明该卷积神经网络的分类准确率高于其他两种方法,最终能在测试集上达到97.2%的准确率,其余两者准确率分别为82.3%、88.7%。结论在少见使用深度卷积神经网络的遥感图像飞机目标分类问题上,本文设计了一个5层卷积神经网络加以应用。实验结果表明,该网络能适应图像场景,自动学习特征,分类效果良好。
[Abstract]:The purpose of the remote sensing image plane target classification, using optical remote sensing image effectively distinguish the type of aircraft, is of great significance to provide military combat information. To solve this problem, there are some traditional machine learning methods, but these methods require manual feature extraction, and difficult to adapt to the complex background of real remote sensing images. In recent years, the rise of depth convolutional neural network method, the network can automatically learn image features and strong generalization ability, is widely used in the field of computer vision. But the depth of convolutional neural network used in classification of remote sensing image plane rare. This paper aims to the depth of the convolutional neural network applied to aircraft remote sensing image classification problems. In the absence of publicly available data sets of the next, a collection of 8 kinds of aircraft data in optical remote sensing images, according to the proportion of 4: 1 roughly divided into training set and test set, and training The training set reasonable expansion. Then according to the particularity of the remote sensing image classification and plane, combined with deep learning convolutional neural network theory, according to the design of a 5 layer convolution neural network. The results of the first training convolutional neural network respectively in the training set gradually extended, and tested using the same test set respectively. Experiments show that, the training set for training network expansion, test accuracy rate increased from 72.4% to 97.2%. in the expanded training set, respectively, to the traditional machine learning method, classical convolutional neural network Le Net-5 and the design of convolutional neural network plan for training, and in the same test set test, the experimental results show that the classification of the convolutional neural network accuracy rate is higher than the other two methods, the final in the test set accuracy rate reached 97.2%, the accuracy rate of the remaining two were 82.3%, 88.7%. conclusion in rare use Based on deep convolution neural network, a 5 level convolution neural network is applied to the classification of aircraft images. The experimental results show that the network can adapt to image scenes, automatically learn features and achieve good classification results.

【作者单位】：北京航空航天大学宇航学院图像中心;航天恒星科技有限公司;
【基金】：国家自然科学基金项目(61671037) 中国航天科技集团公司科技创新研发项目~~
【分类号】：TP183;TP751
【正文快照】： 0引言在现代化战争中,飞机是非常重要的作战工具之一。假如能利用遥感技术获取军事基地的可见光图像,并对机场中的军用飞机进行分类,无疑将带来极大的作战优势。但对遥感图像中的飞机目标进行分类存在背景复杂、飞机数据少、不同型号飞机外形相似等诸多难点。传统的机器学习方

【参考文献】