《基于局部分类器和深度神经网络的遥感图像分类（英文）》由会员分享，可在线阅读，更多相关《基于局部分类器和深度神经网络的遥感图像分类（英文）（10页珍藏版）》请在金锄头文库上搜索。
1、基于局部分类器和深度神经网络的遥感图像分类（英文） 贾厚林 戴红霞 无锡机电高等职业技术学校电子信息工程系 江苏信息职业技术学院电子信息工程学院 摘 要： 深度学习方法作为大数据自动分类工具时表现出较高的性能, 但是在处理遥感图像任务时 (比如图像分类问题) 表现出效率较低。为此, 提出一种新的基于局部分类器和深度神经网络的遥感图像分类算法。首先从原始图像中提取多个局部特征, 并将这些特征输入给用于判决的深度神经网络, 然后按照分配给图像标签对每个局部特征进行分类。最后根据简易的投票方法判决整体图像的结果。利用 World View2 高分辨率卫星遥感影像数据进行了分类实验, 结果显示:提出的
2、方法优于其他分类方法具有较好的分类准确性和分类效率。关键词： 遥感图像分类; 局部分类器; 深度学习; 深度神经网络; 分类性能; 作者简介：Hou-lin JIA, Associate professor. E-mail: 690336783 qq. com收稿日期：7 September 2017Remote sensing image classification based on local partial classifier and deep neural networkHou-lin JIA Hong-xia DAI Department of Electronic and In
3、formation Engineering, Wuxi Machinery and Electron Higher Professional and Technical School; Department of Electronic and Information Engineering, Jiangsu Vocational College of Information Technology; Abstract： The deep learning method exhibits high performance as a large data auto-sorting tool. How
4、ever, when dealing with remote sensing image tasks, such as image classification, the problem of low efficiency is shown.Therefore, a new classification algorithm for remote sensing image based on local classifier and deep neural network is proposed in this paper. First, the method extracts a plural
5、ity of local features from the original image and inputs them into the deep neural network for the decision, and then classifies each local feature according to the assignment to the image tag. Finally, the result of the overall image is judged according to the simple voting method. Based on the Wor
6、ldV iew2 high-resolution satellite remote sensing image data, the classification experiment was carried out. Experimental results show that the proposed method is superior to other classification methods and it has better classification accuracy and classification efficiency.Keyword： Classification
7、of remote sensing images; Local classifier; Deep learning; Deep neural network; Classification performance; Received： 7 September 2017Image classification is a very important task in computer vision tasks, because many social practical applications need the correct image classification results, such
8、 as visual monitoring, marketing, object tracking and large data analysis.Image classification, like other classification problems, is divided into several steps1-3:target detection, image preprocessing, feature extraction and classification.Traditional high resolution remote sensing image classific
9、ation mostly adopts object oriented classification method.During the detection phase, the picture spots of the image will be detected and segmented;Then, the pretreatment technology is used to reduce proportion and illumination difference;The object of picture spots is considered as the basic unit t
10、o realize the classification;However, the classification attribute of object oriented classification method is too high, so the classification accuracy is generally low.It is well known that feature extraction is a key step in achieving good performance.The learning algorithm has become a very hot r
11、esearch field because it could automatically find the best data representation4.The primary purpose of learning algorithms is to automatically convert data into simpler forms so that the useful information can be extracted when the classification is constructed.Deep learning algorithm is a special r
12、epresentation algorithm, which USES neural network to find multi-level features to express abstract concept of data5.These more abstract representations are closer to the semantic content of the data, so they are more useful than raw data and are more conducive to classification.In the computer visi
13、on task, the deep neural network shows its excellent performance.However, using standard deep neural networks to directly study the entire image can be very difficult, especially for complex data such as natural images.Extracting information from the original image sub-region can effectively solve t
14、his problem6.For example, unsupervised learning extracts useful features from image cells, which are called image blocks7.In this paper, the local deep neural network is used to classify the remote sensing image, and a new kind of image classification algorithm based on local feature and deep neural
15、 network is proposed.First, a number of local features are extracted from the original image, and these characteristics are input to the deep neural network for the decision, and then each local feature is classified according to the assignment to the image tag.Finally, all the local features are ta
16、ken into account, using the quick voting method to judge the overall image.The comparison experiment of remote sensing image classification was carried out using the World View2 high-resolution satellite remote sensing image data8.The experimental results show that the classification effect of proposed method is better with higher accuracy and classification efficiency as compared with other classification methods, such as traditional support vector machine and improved
