基于馬爾科夫穩(wěn)態(tài)特性的圖像檢索系統(tǒng)設計與實現(xiàn)
發(fā)布時間：2019-04-11 17:07
【摘要】：隨著計算機、多媒體以及Internet等技術的發(fā)展,尤其是搜索引擎的廣泛應用,人們越來越多的接觸到大量的圖像數(shù)據(jù)。如何快速有效地從大規(guī)模圖像數(shù)據(jù)庫中檢索出所需的圖像已經(jīng)成為目前信息檢索領域非常重要也非常有挑戰(zhàn)性的一個課題。基于內(nèi)容的圖像檢索正是解決這一問題的比較智能且高效的方法。 基于內(nèi)容的圖像檢索方法是根據(jù)圖像中包含物體的類別進行分類的方法,其中圖像特征提取以及分類檢索是最關鍵的兩步,對檢索性能起到關鍵作用,也是近幾年的熱門研究課題�，F(xiàn)有算法比如基于梯度直方圖的圖像特征提取(HOG),金字塔型的梯度直方圖(PHOG)[4]或者SVM分類模型等,盡管已經(jīng)取得了很大的成功,但是基于內(nèi)容的圖像檢索依然是一個很有挑戰(zhàn)性的任務,并且檢索結(jié)果遠不能讓人滿意。原因之一在于圖像的描述符,即圖像的特征提取不充分,另一個原因在于語義相關的圖像在特征空間中,是內(nèi)嵌在一個幾何流形域中,而不是在線性的超平面中。 本課題研究從圖像的特征提取及圖像檢索的快速算法兩個角度出發(fā),在圖像特征提取階段,主要提取的是圖像的梯度信息,并通過馬爾科夫穩(wěn)態(tài)特征的分析,得到該圖像基于馬爾科夫穩(wěn)態(tài)特性的梯度直方圖特征。在圖像分類檢索階段,主要使用了基于幾何流形能量最小化的圖像檢索方法,將圖像的檢索看作一個在圖像數(shù)據(jù)庫中搜索一個最優(yōu)圖像能量環(huán)的問題。實驗結(jié)果表明本文提出的圖像檢索框架是可行的,并且基于馬爾科夫穩(wěn)態(tài)特性的梯度直方圖特征在圖像特征表現(xiàn)上明顯優(yōu)于原始的HOG描述符。 本文的工作主要集中在以下幾點： ●在圖像提取階段,對圖像的梯度直方圖特征進行擴展。根據(jù)馬爾科夫鏈模型,特征化梯度直方圖的空間共生情況,最終通過馬爾科夫穩(wěn)態(tài)特性,得到圖像基于馬爾科夫穩(wěn)態(tài)特性的梯度直方圖特征(GHMSF). GHMSF是對圖像的梯度直方圖特征更進一步的擴展,其中包含了直方圖通道內(nèi)部、直方圖通道與通道之間的空間結(jié)構(gòu)信息。 ●在圖像分類檢索階段是基于幾何流形域,同時將圖像檢索問題看作是圖像數(shù)據(jù)庫中搜索最優(yōu)圖像環(huán)的問題,避免了求解搜索特征空間到語義流形空間的映射關系。 ●禁忌搜索是一個組合優(yōu)化問題,因此挑選最優(yōu)候選解非常耗時。在本文研究中,在搜索過程中采用主動禁忌搜索方法來提高檢索效率。
[Abstract]:With the development of computer, multimedia and Internet technology, especially the extensive application of search engine, more and more people come into contact with a large amount of image data. How to retrieve the required images from large-scale image databases quickly and effectively has become a very important and challenging topic in the field of information retrieval. Content-based image retrieval is an intelligent and efficient method to solve this problem. Content-based image retrieval is a classification method based on the classification of objects in the image. Image feature extraction and classification retrieval are the most important two steps, which play a key role in the retrieval performance. It is also a hot research topic in recent years. Existing algorithms such as gradient histogram-based image feature extraction (HOG), pyramid-type gradient histogram (PHOG) [4] or SVM classification model and so on, although it has achieved great success. However, content-based image retrieval is still a challenging task, and the retrieval results are far from satisfactory. One of the reasons lies in the inadequate feature extraction of the image descriptor, and the other is that the semantic-related image is embedded in a geometric flow field rather than in a linear hyperplane in the feature space, and the other reason is that the semantic-related image is embedded in a geometric flow field rather than in a linear hyperplane. In the stage of image feature extraction, the gradient information of the image is mainly extracted, and through the analysis of Markov steady-state features, this paper studies the feature extraction of image and the fast algorithm of image retrieval from two angles: image feature extraction and fast algorithm of image retrieval. The gradient histogram features of the image based on Markov steady-state characteristics are obtained. In the stage of image classification and retrieval, the image retrieval method based on geometric manifold energy minimization is mainly used, and the image retrieval is regarded as a problem of searching an optimal image energy loop in the image database. The experimental results show that the proposed image retrieval framework is feasible, and the gradient histogram features based on Markov steady-state characteristics are obviously superior to the original HOG descriptors in the performance of image features. The work of this paper mainly focuses on the following points: in the image extraction stage, the gradient histogram features of the image are extended. According to Markov chain model, the spatial symbiosis of gradient histogram is characterized. Finally, the gradient histogram feature (GHMSF). Of image based on Markov steady state is obtained through Markov steady state characteristic. GHMSF is a further extension of the gradient histogram feature of the image, which includes the spatial information between the histogram channel and the histogram channel. In the stage of image classification and retrieval, the problem of image retrieval is based on geometric flow domain. At the same time, the problem of image retrieval is regarded as the problem of searching the optimal image ring in the image database, which avoids the mapping from searching feature space to semantic manifold space. Tabu search is a combinatorial optimization problem, so it is time-consuming to select the optimal solution. In this paper, the active Tabu search method is used to improve the retrieval efficiency.
【學位授予單位】：復旦大學
【學位級別】：碩士
【學位授予年份】：2013
【分類號】：TP391.41
[Abstract]:With the development of computer, multimedia and Internet technology, especially the extensive application of search engine, more and more people come into contact with a large amount of image data. How to retrieve the required images from large-scale image databases quickly and effectively has become a very important and challenging topic in the field of information retrieval. Content-based image retrieval is an intelligent and efficient method to solve this problem. Content-based image retrieval is a classification method based on the classification of objects in the image. Image feature extraction and classification retrieval are the most important two steps, which play a key role in the retrieval performance. It is also a hot research topic in recent years. Existing algorithms such as gradient histogram-based image feature extraction (HOG), pyramid-type gradient histogram (PHOG) [4] or SVM classification model and so on, although it has achieved great success. However, content-based image retrieval is still a challenging task, and the retrieval results are far from satisfactory. One of the reasons lies in the inadequate feature extraction of the image descriptor, and the other is that the semantic-related image is embedded in a geometric flow field rather than in a linear hyperplane in the feature space, and the other reason is that the semantic-related image is embedded in a geometric flow field rather than in a linear hyperplane. In the stage of image feature extraction, the gradient information of the image is mainly extracted, and through the analysis of Markov steady-state features, this paper studies the feature extraction of image and the fast algorithm of image retrieval from two angles: image feature extraction and fast algorithm of image retrieval. The gradient histogram features of the image based on Markov steady-state characteristics are obtained. In the stage of image classification and retrieval, the image retrieval method based on geometric manifold energy minimization is mainly used, and the image retrieval is regarded as a problem of searching an optimal image energy loop in the image database. The experimental results show that the proposed image retrieval framework is feasible, and the gradient histogram features based on Markov steady-state characteristics are obviously superior to the original HOG descriptors in the performance of image features. The work of this paper mainly focuses on the following points: in the image extraction stage, the gradient histogram features of the image are extended. According to Markov chain model, the spatial symbiosis of gradient histogram is characterized. Finally, the gradient histogram feature (GHMSF). Of image based on Markov steady state is obtained through Markov steady state characteristic. GHMSF is a further extension of the gradient histogram feature of the image, which includes the spatial information between the histogram channel and the histogram channel. In the stage of image classification and retrieval, the problem of image retrieval is based on geometric flow domain. At the same time, the problem of image retrieval is regarded as the problem of searching the optimal image ring in the image database, which avoids the mapping from searching feature space to semantic manifold space. Tabu search is a combinatorial optimization problem, so it is time-consuming to select the optimal solution. In this paper, the active Tabu search method is used to improve the retrieval efficiency.
【學位授予單位】：復旦大學
【學位級別】：碩士
【學位授予年份】：2013
【分類號】：TP391.41
【相似文獻】
相關期刊論文 前10條
1 王雙成;冷翠平;劉鳳霞;;無向馬爾科夫毯分類器與集成[J];系統(tǒng)工程與電子技術;2008年07期
2 沈吉鋒;張永志;宋朝河;陳芬;;基于馬爾科夫更新過程的偵察系統(tǒng)可靠性分析[J];兵工自動化;2010年03期
3 馬世榮;;馬爾科夫性是相互獨立性的推廣[J];撫順石油學院學報;1987年02期
4 肖剛;非馬爾科夫可修表決系統(tǒng)可靠性數(shù)字仿真[J];系統(tǒng)工程與電子技術;1998年04期
5 曹建農(nóng),李德仁,關澤群;基于馬爾科夫網(wǎng)視頻圖像目標檢測跟蹤方法研究[J];測繪科學;2004年06期
6 金圣華;周瑋;;馬爾科夫蒙特卡洛在視網(wǎng)膜血管分割中的應用[J];長沙大學學報;2011年05期
7 王輝,王雙成,張劍飛;馬爾科夫網(wǎng)絡中的隱藏變量學習[J];小型微型計算機系統(tǒng);2005年03期
8 曹容菲;張美霞;王醒策;武仲科;周明全;田l，
本文編號：2456600
本文鏈接：http://sikaile.net/kejilunwen/sousuoyinqinglunwen/2456600.html
最近更新
教材專著
