基于多特征的圖像檢索研究
發(fā)布時間：2019-03-23 20:26
【摘要】：圖像數(shù)據(jù)作為互聯(lián)網(wǎng)數(shù)據(jù)中重要的組成部分,隨著互聯(lián)網(wǎng)信息時代的快速發(fā)展以及拍照智能手機的大范圍普及,在以驚人的速度不斷地積累。相比文本數(shù)據(jù),圖像數(shù)據(jù)的優(yōu)勢在于提供了更加豐富和直觀的內(nèi)容資源。那么如何在含有大量圖像的數(shù)據(jù)庫中實現(xiàn)圖像的有效組織和管理,以便人們快速檢索和訪問所需的圖像,這已經(jīng)成為信息時代越來越重要且具有挑戰(zhàn)性的研究問題。最早發(fā)展和建立起來的圖像檢索系統(tǒng)大多都是基于文本檢索方法的,其檢索表現(xiàn)很大程度上依賴于人工標(biāo)注的關(guān)鍵詞信息,而人工標(biāo)注的文本信息既帶有主觀性又無法將圖像中包含的豐富信息完全表達出來。相比于基于文本的圖像檢索方法,傳統(tǒng)的基于內(nèi)容的圖像檢索方法避免了人工標(biāo)注圖像,但是其存在魯棒性不強導(dǎo)致檢索精度不高,或者檢索效率太低等問題,都無法完全滿足實際應(yīng)用的需求。近十年來,在國內(nèi)外的高校、科研機構(gòu)的努力下,局部特征(如SIFT)和視覺詞袋模型(bag-of-visual-words,BoVW)的提出和應(yīng)用,大大推動了基于內(nèi)容的圖像檢索向更高的層次發(fā)展。本文主要研究基于多特征的圖像檢索,提出了一種新的圖像檢索算法。該圖像檢索算法的核心是聯(lián)合圖像語義屬性特征的二維倒排索引,能使那些與查詢圖像有大量相似局部特征且語義相似的候選圖像在檢索結(jié)果中排序更靠前,使得檢索結(jié)果更符合用戶需求。論文介紹了基于內(nèi)容的圖像檢索所涉及到的相關(guān)理論,重點分析了傳統(tǒng)的倒排索引和圖像語義屬性特征的提取,在此基礎(chǔ)上引出了本文研究的主要內(nèi)容:二維倒排索引的構(gòu)建和二維倒排索引的更新。在二維倒排索引的構(gòu)建階段,首先利用高斯差分(DoG)檢測出圖像的尺度不變關(guān)鍵點(Keypoint),基于尺度不變關(guān)鍵點提取SIFT特征和CN(Color Names)顏色特征,然后采用獨立數(shù)據(jù)集上訓(xùn)練的SIFT視覺詞典和CN視覺詞典對兩種局部特征進行量化,隨后圖像的每個關(guān)鍵點都表示為一個視覺單詞對,進而構(gòu)建出二維倒排索引。二維倒排索引相當(dāng)于在索引層面上融合了兩種視覺特征信息,能減少關(guān)鍵點的匹配錯誤。在二維倒排索引的更新階段,本文先將圖像語義屬性特征轉(zhuǎn)變成概率向量,用概率向量的總方差距離(Total Variance Distance,TVD)衡量圖像之間的語義相似度。然后遍歷已構(gòu)建的二維倒排索引,將與圖像的內(nèi)容語義很相似的若干數(shù)據(jù)庫圖像插入該圖像所在的倒排索引項中,如果待插入的圖像已經(jīng)存在于當(dāng)前倒排列表中就不執(zhí)行插入操作。通過這種索引更新方式能在二維倒排索引中聯(lián)合圖像語義屬性特征,使得圖像檢索精度在一定程度上得到提高。本文分別在Ukbench和Holidays圖像數(shù)據(jù)集上對提出的檢索算法進行了實驗驗證,實驗結(jié)果表明本文提出的基于多特征的圖像檢索算法能夠獲得較好的檢索表現(xiàn),在近似重復(fù)圖像檢索中具有一定的應(yīng)用價值。
[Abstract]:Image data as an important part of Internet data, with the rapid development of the Internet information age and the widespread popularity of photo-taking smartphones, it is accumulating at an astonishing speed. Compared with text data, the advantage of image data is that it provides more abundant and intuitive content resources. So how to organize and manage images effectively in the database containing a large number of images so that people can quickly retrieve and access the required images has become a more and more important and challenging research issue in the information age. Most of the earliest developed and established image retrieval systems are based on text retrieval methods, and their retrieval performance depends to a large extent on manually labeled keyword information. The text information of manual annotation is not only subjective but also unable to express the rich information contained in the image. Compared with the text-based image retrieval method, the traditional content-based image retrieval method avoids manual labeling of images, but its robustness is not strong enough to lead to low retrieval accuracy or low retrieval efficiency, and so on. Can not fully meet the needs of practical applications. In the past decade, with the efforts of domestic and foreign universities and scientific research institutions, local features (such as SIFT) and visual word bag model (bag-of-visual-words,BoVW) have been put forward and applied. It greatly promotes the development of content-based image retrieval to a higher level. In this paper, multi-feature-based image retrieval is studied, and a new image retrieval algorithm is proposed. The core of the image retrieval algorithm is the two-dimensional inverted index which combines the semantic attribute features of the image, which can make the candidate images which have a lot of similar local features and similar semantics of the query image rank further in the retrieval results. So that the retrieval results are more in line with the needs of the user. This paper introduces the related theories involved in content-based image retrieval, and focuses on the analysis of the traditional inverted index and the extraction of image semantic attribute features. On this basis, the main contents of this paper are introduced: the construction of two-dimensional inverted index and the updating of two-dimensional inverted index. In the construction stage of two-dimensional inverted index, firstly, Gao Si differential (DoG) is used to detect the scale-invariant key points of the image. (Keypoint), extracts the SIFT features and CN (Color Names) color features based on the scale-invariant keys. Then the SIFT visual dictionary and the CN visual dictionary trained on the independent dataset are used to quantify the two local features. Then each key point of the image is represented as a visual word pair, and then a two-dimensional inverted index is constructed. The two-dimensional inverted index is equivalent to the fusion of two kinds of visual feature information at the index level, which can reduce the matching errors of key points. In the updating stage of two-dimensional inverted index, the semantic attribute feature of image is transformed into probability vector firstly, and the semantic similarity between images is measured by the total variance distance of probability vector (Total Variance Distance,TVD). You then traverse the built two-dimensional inverted index and insert a number of database images that are similar to the content semantics of the image into the inverted index item in which the image is located. Inserts are not performed if the image to be inserted already exists in the current inverted list. Through this index updating method, we can combine semantic attribute features in two-dimensional inverted index, and improve the image retrieval accuracy to a certain extent. The experimental results show that the proposed image retrieval algorithm based on multi-features can achieve better retrieval performance in the Ukbench and Holidays image datasets, and the experimental results show that the proposed image retrieval algorithm based on multi-feature can achieve better retrieval performance. It has certain application value in approximate repeated image retrieval.
【學(xué)位授予單位】：西南大學(xué)
【學(xué)位級別】：碩士
【學(xué)位授予年份】：2017
【分類號】：TP391.41
本文編號：2446198
[Abstract]:Image data as an important part of Internet data, with the rapid development of the Internet information age and the widespread popularity of photo-taking smartphones, it is accumulating at an astonishing speed. Compared with text data, the advantage of image data is that it provides more abundant and intuitive content resources. So how to organize and manage images effectively in the database containing a large number of images so that people can quickly retrieve and access the required images has become a more and more important and challenging research issue in the information age. Most of the earliest developed and established image retrieval systems are based on text retrieval methods, and their retrieval performance depends to a large extent on manually labeled keyword information. The text information of manual annotation is not only subjective but also unable to express the rich information contained in the image. Compared with the text-based image retrieval method, the traditional content-based image retrieval method avoids manual labeling of images, but its robustness is not strong enough to lead to low retrieval accuracy or low retrieval efficiency, and so on. Can not fully meet the needs of practical applications. In the past decade, with the efforts of domestic and foreign universities and scientific research institutions, local features (such as SIFT) and visual word bag model (bag-of-visual-words,BoVW) have been put forward and applied. It greatly promotes the development of content-based image retrieval to a higher level. In this paper, multi-feature-based image retrieval is studied, and a new image retrieval algorithm is proposed. The core of the image retrieval algorithm is the two-dimensional inverted index which combines the semantic attribute features of the image, which can make the candidate images which have a lot of similar local features and similar semantics of the query image rank further in the retrieval results. So that the retrieval results are more in line with the needs of the user. This paper introduces the related theories involved in content-based image retrieval, and focuses on the analysis of the traditional inverted index and the extraction of image semantic attribute features. On this basis, the main contents of this paper are introduced: the construction of two-dimensional inverted index and the updating of two-dimensional inverted index. In the construction stage of two-dimensional inverted index, firstly, Gao Si differential (DoG) is used to detect the scale-invariant key points of the image. (Keypoint), extracts the SIFT features and CN (Color Names) color features based on the scale-invariant keys. Then the SIFT visual dictionary and the CN visual dictionary trained on the independent dataset are used to quantify the two local features. Then each key point of the image is represented as a visual word pair, and then a two-dimensional inverted index is constructed. The two-dimensional inverted index is equivalent to the fusion of two kinds of visual feature information at the index level, which can reduce the matching errors of key points. In the updating stage of two-dimensional inverted index, the semantic attribute feature of image is transformed into probability vector firstly, and the semantic similarity between images is measured by the total variance distance of probability vector (Total Variance Distance,TVD). You then traverse the built two-dimensional inverted index and insert a number of database images that are similar to the content semantics of the image into the inverted index item in which the image is located. Inserts are not performed if the image to be inserted already exists in the current inverted list. Through this index updating method, we can combine semantic attribute features in two-dimensional inverted index, and improve the image retrieval accuracy to a certain extent. The experimental results show that the proposed image retrieval algorithm based on multi-features can achieve better retrieval performance in the Ukbench and Holidays image datasets, and the experimental results show that the proposed image retrieval algorithm based on multi-feature can achieve better retrieval performance. It has certain application value in approximate repeated image retrieval.
【學(xué)位授予單位】：西南大學(xué)
【學(xué)位級別】：碩士
【學(xué)位授予年份】：2017
【分類號】：TP391.41
【參考文獻】
相關(guān)期刊論文 前2條
1 羅永興,于明,陳雷;基于內(nèi)容的圖像檢索系統(tǒng)研究[J];計算機與數(shù)字工程;2004年06期
2 陸麗珍,劉仁義,劉南;一種融合顏色和紋理特征的遙感圖像檢索方法[J];中國圖象圖形學(xué)報;2004年03期
相關(guān)碩士學(xué)位論文 前4條
1 林銘德;基于內(nèi)容的圖像數(shù)據(jù)庫檢索技術(shù)的研究及應(yīng)用[D];中南大學(xué);2008年
2 徐慶;基于內(nèi)容的圖像檢索關(guān)鍵技術(shù)研究[D];西安電子科技大學(xué);2008年
3 鄭莉霞;基于文本的Web圖像檢索技術(shù)研究[D];廣西大學(xué);2007年
4 黃春木;基于內(nèi)容的二值圖像檢索技術(shù)研究[D];解放軍信息工程大學(xué);2007年
，本文編號：2446198
本文鏈接：http://sikaile.net/kejilunwen/ruanjiangongchenglunwen/2446198.html
最近更新
教材專著
