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%。接着，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LeNet-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余两者，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。实验验证该网络能适应图像场景，自动学习特征，分类效果良好。
Aircraft classification in remote sensing images using convolutional neural networks
Zhou Min,Shi Zhenwei,Ding Huoping(Image Processing Center, School of Astronautics, Beihang University;Space star technology co,LTD,Beijing)
Objective: Aircraft classification in remote sensing images, in other words, identifying the types of aircraft in remote sensing images rapidly and accurately, is undoubtedly helpful for providing military information and taking military advantages. It is extremely potential to research in the field of processing optical remote sensing images. As for now, there are many conventional machine learning methods for solving this problem. But it is hard to apply these methods to real optical remote sensing images because of the complicated background. What is more, it needs selecting features when using these methods and the performance of classification greatly depends on what features are extracted. So, to gain a relative good result, it usually requires selecting features artificially which is time-consuming and complicated. In recent years, deep convolutional neural network has been very popular. It can learn features by itself and shows excellent generalization ability. And it has been widely used in the areas of computer vision and pattern recognition. However, at present, it is rare to apply convolutional neural networks to this issue. This paper aims at solving aircraft classification problem in optical remote sensing images using convolutional neural networks. Method: Because of lacking public dataset of aircraft in optical remote sensing images, this paper collects eight types of aircraft from optical remote sensing images to form a dataset. There are eight different types of aircraft are collected in this dataset, including boomers, carriers, fighters, primary trainers and tankers. The number of each kind of aircraft are equal. Then the dataset is divided into training set and testing set with the ratio between the size of these two is 4:1. The samples are randomly selected into training dataset or test dataset. But for each type of aircraft, the ratio between the number of training and test dataset stays the same. What is more, as the need of convolutional neural network for data is large, the training set is largely augmented with 3 kinds of methods when training. The final training set is as large as thirty-two times of the original training set. Then, based on the theory of deep convolutional neural network, for aircraft classification in optical remote sensing images, this paper designs a special 5-layer convolutional neural network. To adapt to the speciality of aircraft classification in remote sensing images, such as the lack of data, the low absolute resolution and so on, the size of convolutional kernels and pooling kernels used in shallow layers are both as small as 3×3 pixels. Result: At first, this convolutional neural network is trained respectively on different training sets, which are augmented in different varying degree. Then the well-trained convolutional neural networks are tested using the same test set. The result shows that the test accuracy can be improved from 72.4% to 97.2% through augmenting the training set. Secondly, this paper trains and tests aforementioned 5-layer convolutional neural network on the aircraft dataset. In addition, to guarantee the necessary of data augmentation, this convolutional neural network is respectively trained on original and augmented training dataset. Then the gained networks work on the same test dataset. According to the aforementioned experiment result, the network trained on the largest training dataset performs best. So, the data augmentation is useful and necessary and the dataset mentioned in the following experiment is the augmented one. After establishing the dataset, for better verify the feasibility of convolutional neural network, this paper chooses one kind of classical conventional machine learning method to compare with it. What is more, LeNet-5, the first utilitarian convolutional neural network, is used to identify these eight types of aircraft for comparison too. When in training and testing phase, these three method use the same dataset. The experimental result shows that the classification accuracy achieves as high as 97.2% with this designed 5-layer convolutional neural network. While the accuracies with the other two algorithms are 82.3% and 88.7% respectively. That is to say that this designed 5-layer convolutional neural network performs better than that classical conventional machine learning method and LeNet-5. That is because the conventional machine learning method needs segmenting the images, which is hard to do in complicated scenes. And as for LeNet-5, it is initially for digital numbers recognition rather than aircraft classification. That network is relative smaller and not quite suitable for this issue. Conclusion: This paper designs a 5-layer convolutional neural network especially for aircraft target classification in optical remote sensing images, in which deep convolutional neural network is rare to see. The experiments show that this convolutional neural network can learn the features of aircraft well and sort these aircraft with a high classification accuracy.