深度学习在高光谱图像的降维及分类中的应用

发布时间：2018-04-20 00:41

本文选题：高光谱图像 + 卷积神经网络　；参考：《电子科技大学》2017年硕士论文

【摘要】：遥感技术的发展使得新的遥感传感器可以采集的图像具有连续的谱域和空域,这些图像含有大量的地物信息——地物的光谱信息和几何空间分布。传统的遥感图像分类框架只利用了谱域信息来分类,忽略了空域信息对分类的影响。卷积神经网络有着独特的优势。图像不需要太多的前期处理就可以直接输入网络,从训练数据中隐式地进行学习,规避了特征提取和分类中数据重建的过程。它独特的层间联系以及空间信息的密切联系,使其适用于图像处理中分类识别任务。本文在充分考虑高光谱图像特点的前提下,提出了在图像分类领域中取得显著成果的卷积神经网络来对高光谱图像的像元进行分类。本文主要研究内容:(1)借鉴LeNet-5网络框架的设计思想,设计适用于高光谱图像分类的卷积神经网络框架。本文主要研究了网络框架中的层数、卷积层中神经元的数目、下采样层中神经元的数目以及输出层中神经元的数目,达到了高光谱数据分类到图像分类的有效转化。(2)将每个像素点的空间邻域信息作为卷积神经网络框架的输入样本,探究所设计的框架对高光谱图像分类的有效性。(3)研究框架中的激活函数ReLU的设计,达到缓解梯度弥散目的,提高网络的执行效率和分类精度。与梯度下降法相比,mini-batch随机梯度下降法可以大大提高框架的执行效率。这些策略的使用更有助于提取分类特征和提高分类效果。(4)利用本文所设计的框架对The University of Pavia数据集进行分类仿真实验,验证其可行性,并与传统的k近邻、BP神经网络以及SVM分类识别方法比较。实验仿真结果表明:采用本文所设计的分类框架,分类精度高于其它分类方法,达到了97.57%。
[Abstract]:With the development of remote sensing technology, the images which can be collected by the new remote sensing sensor have continuous spectral domain and spatial domain. These images contain a lot of ground object information-spectral information and geometric spatial distribution. The traditional remote sensing image classification framework only uses spectral domain information to classify, ignoring the influence of spatial information on classification. Convolutional neural networks have unique advantages. The image can be directly input into the network without too much pre-processing, and can be learned implicitly from the training data, thus avoiding the process of feature extraction and data reconstruction in classification. It is suitable for classification and recognition in image processing because of its unique interlayer connection and close connection of spatial information. On the premise of fully considering the characteristics of hyperspectral images, a convolution neural network, which has achieved remarkable results in the field of image classification, is proposed in this paper to classify the pixels of hyperspectral images. In this paper, we use the design idea of LeNet-5 network framework for reference, and design a convolution neural network framework suitable for hyperspectral image classification. In this paper, the number of layers in the network framework, the number of neurons in the convolution layer, the number of neurons in the lower sampling layer and the number of neurons in the output layer are studied. An effective transformation from hyperspectral data classification to image classification is achieved. The spatial neighborhood information of each pixel is used as the input sample of the convolution neural network framework. The effectiveness of the proposed framework for hyperspectral image classification is explored. (3) the design of the activation function (ReLU) in the framework is studied in order to alleviate the gradient dispersion and improve the efficiency and classification accuracy of the network. Compared with the gradient descent method, the mini-batch stochastic gradient descent method can greatly improve the performance efficiency of the frame. The use of these strategies is more helpful to extract the classification features and improve the classification effect. (4) using the framework designed in this paper, the The University of Pavia dataset classification simulation experiments are carried out to verify its feasibility. And compared with the traditional k-nearest neighbor BP neural network and SVM classification and recognition methods. The experimental results show that the classification accuracy of the proposed classification framework is higher than that of other classification methods, reaching 97.57.
【学位授予单位】：电子科技大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP751

【参考文献】