結合特徵線轉換法的生成對抗網路模型;Generative Adversarial Net Based on Feature Line Embedding

NCUIR > Research Center > Center for Space and Remote Sensing Research > Research Project > Item 987654321/82073

Please use this identifier to cite or link to this item: http://ir.lib.ncu.edu.tw/handle/987654321/82073

Title:	結合特徵線轉換法的生成對抗網路模型;Generative Adversarial Net Based on Feature Line Embedding
Authors:	陳映濃
Contributors:	國立中央大學太空及遙測研究中心
Keywords:	遙測影像;特徵空間;拓樸;生成對抗網路;remote sensing image;feature space;topology;Generative Adversarial Net
Date:	2020-01-13
Issue Date:	2020-01-13 14:13:53 (UTC+8)
Publisher:	科技部
Abstract:	遙測影像的應用一直以來是熱門的研究議題，例如從紅外線衛星影像(Infrared Image, IR)中進行土地分類便是一項重要的工作。然而，由於IR影像對影像處理來說是一種高維影像，如何從中擷取對土地分類有用的資訊便是一項重要的工作。因此，對高維的衛星影像進行降維，並且在降維空間中擷取重要特徵便直接關係著後續分類結果的成效。除此之外，降維工具除了能夠擷取重要特徵之外，還要能夠保持原資料的拓樸特性，若能保持好原資料的拓樸特性，其實也就增強了特徵空間的區別能力。有鑑於此，我們在本計畫中提出了Fuzzy Kernel Feature Line Embedding，它可以將資料在原空間中的非線性、非尤拉距離的資訊保留下來，並同時可以保留原空間中資料的拓樸特性。而資料在原空間中的拓樸特性即我們認為是一項生成資料良是否足夠擬真的重要指標，因此我們將GAN中的對抗網路判別過程放在FKFLE所建立的特徵空間中來進行，如此即可以判別生成資料是否俱備與原資料相同的拓樸架構，有了相同的拓樸架構，即俱備了相當足夠的一般性，因此可以用來補足遙測影像中訓練資料不足的狀況。 ;The application of remote sensing images has always been a hot research topic. For example, land classification from Infrared Image (IR) is an important task. However, since IR images are high-dimensional images for image processing, how to extract useful information for land classification is an important task. Therefore, the dimension reduction of high-dimensional satellite images and the acquisition of important features in the dimension reduction space are directly related to the effectiveness of subsequent classification results. In addition, in addition to the important features of the dimension reduction tool, it is also necessary to maintain the topological characteristics of the original data. If the topology of the original data is maintained, the ability to distinguish the feature space is enhanced. In view of this, we proposed the Fuzzy Kernel Feature Line Embedding in this project, which can preserve the information of nonlinear and non-Euclidean distance in the original space, and at the same time preserve the topology of the data in the original space. . The topology of the data in the original space is considered to be an important indicator of whether the generated data is good enough. Therefore, we put the anti-network discriminating process in GAN into the feature space established by FKFLE. That is to say, it can be determined whether the generated data is the same as the original data. With the same topology, it is quite sufficient generality, so it can be used to make up for the lack of training data in telemetry images.
Relation:	財團法人國家實驗研究院科技政策研究與資訊中心
Appears in Collections:	[Center for Space and Remote Sensing Research ] Research Project

Files in This Item:

File	Description	Size	Format
index.html		0Kb	HTML	265	View/Open

社群 sharing

Loading...