基于詞袋模型的圖像分類技術(shù)研究
發(fā)布時(shí)間：2018-08-27 11:02
【摘要】：圖像分類是數(shù)字圖像分析與理解的主要研究方向,因而受到人們廣泛的關(guān)注,如何對(duì)大量的圖像進(jìn)行快速、準(zhǔn)確的分類是當(dāng)前研究中的一個(gè)熱點(diǎn)問題。 本文將詞袋模型應(yīng)用于無監(jiān)督圖像分類,并針對(duì)傳統(tǒng)詞袋模型中所涉及方法(如特征提取、聚類方法等)的不足,做出了如下的改進(jìn)： 首先,采用加速魯棒特征變換(Speeded Up Robust Features, SURF)提取特征(圖像詞語)。本文將增加預(yù)處理環(huán)節(jié),即通過特征興趣區(qū)域(Region of interesting Features, ROIF)與圖像前景對(duì)象范圍定位方法對(duì)特征提取范圍進(jìn)行規(guī)定。實(shí)驗(yàn)表明,預(yù)處理環(huán)節(jié)可有效減少特征中弱特征與無關(guān)特征數(shù)量。 其次,本文采用精確歐拉位置敏感哈希(Exact Euclidean Locality Sensitive Hashing, E2LSH)對(duì)圖像詞語聚類以獲得關(guān)鍵詞(聚類中心)。為了減小E2LSH的隨機(jī)性,將多次聚類并使用基于最短特征無向圖的組合匯集技術(shù)獲得最終的聚類分布。實(shí)驗(yàn)表明,由E2LSH得到的關(guān)鍵詞具有更強(qiáng)的代表性。 再次,利用吉布斯抽樣計(jì)算隱狄利克雷分布模型(Latent Dirichlet Allocation, LDA)得到類別分布轉(zhuǎn)移矩陣。并使用最大轉(zhuǎn)移概率與轉(zhuǎn)移向量相似度結(jié)合的組合方法閱讀轉(zhuǎn)移矩陣獲得分類結(jié)果。實(shí)驗(yàn)表明,組合式閱讀能更好的發(fā)現(xiàn)轉(zhuǎn)移矩陣中隱含的類別信息。 最后,針對(duì)傳統(tǒng)分類方法對(duì)同一類內(nèi)圖像間關(guān)系的忽視,利用雙向匹配(Bidirectional Matching, BM),隨機(jī)抽樣一致性(Random Sample Consensus, RANSAC)和感知哈希(Perceptual Hash, pHash)進(jìn)行同一類內(nèi)圖像間關(guān)系的尋找(特征匹配)。實(shí)驗(yàn)表明,通過上述方法可準(zhǔn)確獲得同一類內(nèi)圖像間的關(guān)系,實(shí)現(xiàn)類內(nèi)圖像的細(xì)化分類。
[Abstract]:Image classification is the main research direction of digital image analysis and understanding, so people pay more attention to it. How to classify a large number of images quickly and accurately is a hot issue in current research. This paper applies the word bag model to unsupervised image classification, and aiming at the shortcomings of the traditional word bag model (such as feature extraction, clustering method, etc.), the following improvements are made: first, An accelerated robust feature transform (Speeded Up Robust Features, SURF) is used to extract features (image words). In this paper, preprocessing is added, that is, the range of feature extraction is defined by the region of interest (Region of interesting Features, ROIF) and the image foreground object localization method. Experiments show that preprocessing can effectively reduce the number of weak and independent features. Secondly, the accurate Euler position sensitive (Exact Euclidean Locality Sensitive Hashing, E2LSH (Euler Hash (Exact Euclidean Locality Sensitive Hashing, E2LSH) is used to cluster the image words to obtain the key words (clustering center). In order to reduce the randomness of E2LSH, the final clustering distribution is obtained by using the combination aggregation technique based on the shortest feature undirected graph. Experiments show that the keywords obtained from E2LSH are more representative. Thirdly, the class distribution transfer matrix is obtained by using Gibbs sampling to calculate the hidden Dirichlet distribution model (Latent Dirichlet Allocation, LDA). A combination of maximum transition probability and transfer vector similarity is used to read the transition matrix to obtain the classification results. The experimental results show that the combined reading can better find the category information implied in the transfer matrix. Finally, aiming at the neglect of the relationship between images in the same class by the traditional classification methods, we use two-way matching (Bidirectional Matching, BM), random sampling consistent (Random Sample Consensus, RANSAC) and perceptual hash (Perceptual Hash, pHash) to find the relationship (feature matching) between images within the same class. The experimental results show that the relationship between the images in the same class can be accurately obtained by the above methods, and the thinning and classification of the images in the same class can be realized.
【學(xué)位授予單位】：安徽大學(xué)
【學(xué)位級(jí)別】：碩士
【學(xué)位授予年份】：2014
【分類號(hào)】：TP391.41
[Abstract]:Image classification is the main research direction of digital image analysis and understanding, so people pay more attention to it. How to classify a large number of images quickly and accurately is a hot issue in current research. This paper applies the word bag model to unsupervised image classification, and aiming at the shortcomings of the traditional word bag model (such as feature extraction, clustering method, etc.), the following improvements are made: first, An accelerated robust feature transform (Speeded Up Robust Features, SURF) is used to extract features (image words). In this paper, preprocessing is added, that is, the range of feature extraction is defined by the region of interest (Region of interesting Features, ROIF) and the image foreground object localization method. Experiments show that preprocessing can effectively reduce the number of weak and independent features. Secondly, the accurate Euler position sensitive (Exact Euclidean Locality Sensitive Hashing, E2LSH (Euler Hash (Exact Euclidean Locality Sensitive Hashing, E2LSH) is used to cluster the image words to obtain the key words (clustering center). In order to reduce the randomness of E2LSH, the final clustering distribution is obtained by using the combination aggregation technique based on the shortest feature undirected graph. Experiments show that the keywords obtained from E2LSH are more representative. Thirdly, the class distribution transfer matrix is obtained by using Gibbs sampling to calculate the hidden Dirichlet distribution model (Latent Dirichlet Allocation, LDA). A combination of maximum transition probability and transfer vector similarity is used to read the transition matrix to obtain the classification results. The experimental results show that the combined reading can better find the category information implied in the transfer matrix. Finally, aiming at the neglect of the relationship between images in the same class by the traditional classification methods, we use two-way matching (Bidirectional Matching, BM), random sampling consistent (Random Sample Consensus, RANSAC) and perceptual hash (Perceptual Hash, pHash) to find the relationship (feature matching) between images within the same class. The experimental results show that the relationship between the images in the same class can be accurately obtained by the above methods, and the thinning and classification of the images in the same class can be realized.
【學(xué)位授予單位】：安徽大學(xué)
【學(xué)位級(jí)別】：碩士
【學(xué)位授予年份】：2014
【分類號(hào)】：TP391.41
【參考文獻(xiàn)】
相關(guān)期刊論文 前7條
1 茹朝陽;;基于SIFT特征的遙感影像自動(dòng)配準(zhǔn)[J];城市勘測(cè);2009年06期
2 趙永威;李弼程;彭天強(qiáng);高毫林;;一種基于隨機(jī)化視覺詞典組和查詢擴(kuò)展的目標(biāo)檢索方法[J];電子與信息學(xué)報(bào);2012年05期
3 騫森;朱劍英;;基于改進(jìn)的SIFT特征的圖像雙向匹配算法[J];機(jī)械科學(xué)與技術(shù);2007年09期
4 嚴(yán)華;;一種改進(jìn)的k-means算法[J];計(jì)算機(jī)與現(xiàn)代化;2009年01期
5 石晶;范猛;李萬龍;;基于LDA模型的主題分析[J];自動(dòng)化學(xué)報(bào);2009年12期
6 何周燦;王慶;楊恒;;一種面向快速圖像匹配的擴(kuò)展LSH算法[J];四川大學(xué)學(xué)報(bào)(自然科學(xué)版);2010年02期
7 劉守群;朱明;鄭p，
本文編號(hào)：2207077
本文鏈接：http://sikaile.net/kejilunwen/ruanjiangongchenglunwen/2207077.html
最近更新
教材專著
