本文關(guān)鍵詞： 半監(jiān)督分類 高光譜圖像 DL1圖 KNN圖 標(biāo)記傳播　出處：《北方民族大學(xué)》2017年碩士論文　論文類型：學(xué)位論文

【摘要】：當(dāng)前,機(jī)器學(xué)習(xí)的相關(guān)理論和應(yīng)用研究遍地開花。傳統(tǒng)機(jī)器學(xué)習(xí)常用的兩種方法為無監(jiān)督學(xué)習(xí)和有監(jiān)督學(xué)習(xí)。然而我們也應(yīng)該看到,無監(jiān)督學(xué)習(xí)的特點(diǎn)和優(yōu)勢是不需要訓(xùn)練樣本,但無監(jiān)督學(xué)習(xí)對于空間分布較復(fù)雜的數(shù)據(jù)難以得到好的學(xué)習(xí)效果;另一方面,有監(jiān)督學(xué)習(xí)雖然學(xué)習(xí)性能較好,但當(dāng)訓(xùn)練樣本數(shù)目很少時,有監(jiān)督學(xué)習(xí)難以準(zhǔn)確學(xué)習(xí)出樣本的真實(shí)分布。由于半監(jiān)督學(xué)習(xí)可以利用大量的未標(biāo)記樣本來輔助有限的有標(biāo)記樣本,從而避免無監(jiān)督學(xué)習(xí)和有監(jiān)督學(xué)習(xí)兩種方法的弊端,結(jié)合其優(yōu)勢,因而半監(jiān)督學(xué)習(xí)可以提高學(xué)習(xí)的準(zhǔn)確性。由于半監(jiān)督學(xué)習(xí)可以在得到較高分類精度的同時降低學(xué)習(xí)的成本,它對在理論和實(shí)際中提高學(xué)習(xí)器的分類性能有著非常重要的指導(dǎo)意義。在半監(jiān)督學(xué)習(xí)的眾多方法中,基于圖論的半監(jiān)督學(xué)習(xí)方法在實(shí)踐中取得了實(shí)證性的成功,它成為了半監(jiān)督學(xué)習(xí)方法中最流行的一種。但是,基于圖的半監(jiān)督學(xué)習(xí)方法有其特殊性,構(gòu)圖方法會對學(xué)習(xí)器的性能產(chǎn)生較大影響。針對構(gòu)圖問題,我們提出一種基于可區(qū)分L1范數(shù)圖和KNN圖疊加圖的半監(jiān)督分類算法。并將該算法用于小樣本高光譜圖像分類問題上。本文的主要研究工作如下:基于少數(shù)已標(biāo)記樣本的高光譜圖像分類是一個具有挑戰(zhàn)的任務(wù)。對于基于圖的方法,如何構(gòu)造圖是分類成功的關(guān)鍵。本文提出一種新的構(gòu)圖方法,首先在L1范數(shù)圖基礎(chǔ)上構(gòu)造一個區(qū)分能力更強(qiáng)的L1圖,即DL1圖,并將其和KNN圖疊加,在半監(jiān)督框架下來解決高光譜圖像的分類問題。本文的圖構(gòu)造方法包括兩個步驟,首先,使用稀疏表示方法估計(jì)任意兩個像素屬于同一類別的概率,構(gòu)建相應(yīng)概率矩陣,然后再將概率矩陣整合到L1圖中,從而得到DL1圖。其次,將DL1圖與KNN圖以一定比例線性疊加。在Indiana Pines高光譜數(shù)據(jù)集上的實(shí)驗(yàn)表明,所提方法的分類識別率更高。本文將概率矩陣與L1圖的權(quán)值矩陣疊加,形成了強(qiáng)鑒別力的DL1圖。將空間的局部信息與光譜的全局信息通過KNN圖和DL1圖結(jié)合在一起,構(gòu)建了空譜信息聯(lián)合的圖框架結(jié)構(gòu),使用該框架構(gòu)建的圖,能更精細(xì)的反映高光譜圖像數(shù)據(jù)的圖譜結(jié)構(gòu)。利用圖的標(biāo)記傳播達(dá)到半監(jiān)督分類的目的,以此提高小樣本高光譜圖像自動分類的精度,實(shí)驗(yàn)表明,在標(biāo)記樣本比例為5%時,分類精度提升亦非常顯著。
[Abstract]:At present, machine learning theory and Application Research of traditional machine learning. Blossom everywhere the two commonly used methods for unsupervised learning and supervised learning. However, we should also see that without the characteristics and advantages of supervised learning is not required for the training samples, but unsupervised learning for the distribution of data is complex it is difficult to get good learning effect; on the other hand, supervised learning is better learning performance, but when the number of training samples is small, supervised learning is difficult to learn the true distribution of the sample. The semi supervised learning can use unlabeled samples to assist limited labeled samples, so as to avoid the disadvantages of unsupervised learning and supervised learning two methods, combined with its advantages, so the semi supervised learning can improve the learning accuracy. Because semi supervised learning can get higher classification precision and lower The cost of learning, it has a very important significance to improve the classification performance of learning in theory and practice. Many methods in the semi supervised learning, semi supervised learning method based on graph theory has achieved substantial success in practice, it has become one of the most popular methods in the semi supervised learning however, graph based semi supervised learning method has its particularity, patterning method will have great influence on the performance of classifier. Aiming at the problem of composition, we propose a discriminative semi supervised classification algorithm based on L1 norm and KNN diagram based on superposition graph. And the algorithm for hyperspectral image classification problems. The main research work of this paper is as follows: hyperspectral image classification based on few labeled samples is a challenging task for graph based methods, how to construct a map is the key to success. This paper proposes a classification A new method of composition, a stronger ability to distinguish L1 graph constructed in L1 norm map based on DL1 map and KNN map, and the overlay, solve the classification problem of hyperspectral image in the semi supervised framework. The graph construction method includes two steps: firstly, using sparse representation method the estimated probability of any two pixels belonging to the same category, construct the corresponding probability matrix, and then the probability matrix is integrated into the L1 map, DL1 map is obtained. Secondly, the DL1 map and KNN map to a certain proportion of linear superposition. Show in Indiana Pines hyperspectral data set on the experiment, the proposed method of classification and recognition the rate is high. The weight matrix superposition probability matrix and L1 map, the formation of a strong discrimination DL1. The global information and local information of spectral space by KNN map and DL1 map together, constructs the space spectrum framework of information combination The structure, the use of the framework construction of the map, which can reflect the structure of hyperspectral image data is more accurate. To achieve semi supervised classification using marker propagation graph for the purpose, in order to improve the classification accuracy of hyperspectral images of small samples, experiments show that, in the proportion of labeled samples is 5%, enhance the classification accuracy is also very significant.

【學(xué)位授予單位】：北方民族大學(xué)
【學(xué)位級別】：碩士
【學(xué)位授予年份】：2017
【分類號】：TP391.41;TP181

【參考文獻(xiàn)】

相關(guān)期刊論文 前10條

1 張?jiān)票?張春梅;周千琪;戴模;;基于l~1范數(shù)和k近鄰疊加圖的半監(jiān)督分類算法[J];模式識別與人工智能;2016年09期

2 樊利恒;呂俊偉;鄧江生;;基于分類器集成的高光譜遙感圖像分類方法[J];光學(xué)學(xué)報(bào);2014年09期

3 馬爾侖;鄭艷楠;;基于光譜特征的監(jiān)督分類方法在黃河口濕地的應(yīng)用比較[J];價值工程;2014年02期

4 宋相法;焦李成;;基于稀疏表示及光譜信息的高光譜遙感圖像分類[J];電子與信息學(xué)報(bào);2012年02期

5 李民;程建;李小文;樂翔;;非局部學(xué)習(xí)字典的圖像修復(fù)[J];電子與信息學(xué)報(bào);2011年11期

6 王雪松;張曉麗;程玉虎;;一種簡潔局部全局一致性學(xué)習(xí)[J];控制與決策;2011年11期

7 錢育蓉;賈振紅;于炯;楊峰;段文亮;;BP-ANN在荒漠草地高光譜分類研究中的應(yīng)用[J];計(jì)算機(jī)工程與應(yīng)用;2011年12期

8 梁吉業(yè);高嘉偉;常瑜;;半監(jiān)督學(xué)習(xí)研究進(jìn)展[J];山西大學(xué)學(xué)報(bào)(自然科學(xué)版);2009年04期

9 孫玉寶;肖亮;韋志輝;邵文澤;;基于Gabor感知多成份字典的圖像稀疏表示算法研究[J];自動化學(xué)報(bào);2008年11期

10 韓建峰;楊哲海;;組合分類器及其在高光譜影像分類中的應(yīng)用[J];測繪科學(xué)技術(shù)學(xué)報(bào);2007年03期

相關(guān)博士學(xué)位論文 前1條

1 劉春紅;超光譜遙感圖像降維及分類方法研究[D];哈爾濱工程大學(xué);2005年

，

本文編號：1547826