本文關鍵詞： 天體測量學 技術 圖像處理 方法 數(shù)據(jù)分析 方法 統(tǒng)計　出處：《天文學報》2017年04期 　論文類型：期刊論文

【摘要】：Cassini的光學成像系統(tǒng)(Imaging Science Subsystem,ISS)拍攝了大量的土星及其衛(wèi)星的圖像,其中一部分可以用來做天體測量工作,但是需要人工挑揀出來,這是一項繁重的工作.研究目的是將這種工作自動化.為此,將卷積神經(jīng)網(wǎng)絡(Convolution Neural Network,CNN)與支持向量機(Support Vector Machine,SVM)結合起來,提出了一種ISS圖像可測性分類系統(tǒng).系統(tǒng)首先通過深度卷積網(wǎng)絡提取ISS圖像的特征描述子,然后使用SVM分類器根據(jù)圖像的特征描述子對圖像進行分類.對比了3種有代表性的深度卷積網(wǎng)絡:CNN-F、CNN-M-128和Very Deep-19,實驗結果表明:CNN-F卷積網(wǎng)絡加SVM可以提供較好的分類結果,其分類準確率在97%以上.研究不僅可用于Cassini ISS圖像的天體測量工作,也可以推廣到其他空間探測項目的類似工作中.
[Abstract]:Cassini's optical imaging system, Imaging Science Subsystem ISS, takes a lot of images of Saturn and its moons, some of which can be used for astrometry, but they need to be picked up manually. This is a heavy task. The purpose of the study is to automate this work. To do this, we combine the convolutional neural network solution Neural Network CNNs with the support Vector Machine (SVM). In this paper, a testability classification system for ISS image is proposed. Firstly, the feature descriptor of ISS image is extracted by deep convolution network. Then we use SVM classifier to classify the images according to the feature descriptors of the image. Three typical deep convolution networks: CNN-FU CNN-M-128 and Very Deep-19 are compared. The experimental results show that the SVM plus the convolution network of the SVM can provide better classification results. The classification accuracy is more than 97%. The research can be used not only in the astrometry of Cassini ISS images, but also in the similar work of other space exploration projects.
【作者單位】： 暨南大學計算機系;
【基金】：國家自然科學基金委員會-中國科學院天文聯(lián)合基金重點項目(U1431227) 國家自然科學基金項目(11403008) 廣東省自然科學基金項目(2016A030313092、2014A030313374) 廣東省教育廳高等學�？萍紕�(chuàng)新項目(2013KJCX0020)資助
【分類號】：P12
，

本文編號：1555581