基于三重低秩正則化的魯棒半監(jiān)督多標(biāo)記學(xué)習(xí)算法及其在圖像標(biāo)注中的應(yīng)用
發(fā)布時(shí)間:2021-04-27 18:11
圖像自動(dòng)標(biāo)注技術(shù)在圖像檢索領(lǐng)域發(fā)揮著越來(lái)越重要的作用,逐漸成為計(jì)算機(jī)視覺(jué)的研究熱點(diǎn)。數(shù)字可視化技術(shù)的進(jìn)步和發(fā)展使得大量的圖像可以在網(wǎng)絡(luò)上獲取,用戶(hù)可以根據(jù)自己的喜好從存儲(chǔ)庫(kù)中檢索這些圖像,然而這些圖像大多數(shù)都沒(méi)有描述信息。圖像標(biāo)注的傳統(tǒng)做法是由人類(lèi)來(lái)完成的,這是一種費(fèi)時(shí)費(fèi)力的標(biāo)注方法,也是一種過(guò)于主觀的標(biāo)注方法。另一個(gè)難點(diǎn)是解決低層視覺(jué)特征(顏色、形狀和紋理)與用于解釋圖像的高層語(yǔ)義特征之間的語(yǔ)義鴻溝問(wèn)題。大多數(shù)圖像檢索方法是基于內(nèi)容的圖像檢索(CBIR)和基于標(biāo)簽的圖像檢索(TBIR)方法。CBIR通過(guò)提取圖像本身的顏色、紋理和形狀等特征,在低層特征上進(jìn)行工作,但由于語(yǔ)義鴻溝,一般用戶(hù)無(wú)法使用它。TBIR的工作原理是根據(jù)文本查詢(xún)和圖像的手工標(biāo)注之間的匹配來(lái)查找相關(guān)圖像。但是它高度依賴(lài)于標(biāo)簽的可用性和質(zhì)量。然而,手動(dòng)標(biāo)注的標(biāo)記是主觀的、模糊的、有限且?guī)г肼暤摹=陙?lái),該領(lǐng)域的研究已經(jīng)通過(guò)圖像自動(dòng)標(biāo)注的方法(AIA),將低層圖像特征與高層語(yǔ)義之間的語(yǔ)義鴻溝聯(lián)系起來(lái)。自動(dòng)圖像標(biāo)注算法假設(shè)采集的圖像樣本具有語(yǔ)義標(biāo)記和低層特征表示。該標(biāo)注方法使用機(jī)器學(xué)習(xí)算法,然后可以訓(xùn)練它使用低層特征進(jìn)行語(yǔ)義...
【文章來(lái)源】:北京交通大學(xué)北京市 211工程院校 教育部直屬院校
【文章頁(yè)數(shù)】:68 頁(yè)
【學(xué)位級(jí)別】:碩士
【文章目錄】:
致謝
摘要
Abstract
Chapter 1 Introduction
1.1 Background
1.2 Problem of statement
1.3 Scope and objectives
1.4 Methodology
1.5 Thesis outline
Chapter 2 Literature Review
2.1 Background
2.2 Automatic Image annotation
2.3 Multi-Label Learning
2.4 Semi-Supervised Multi-Label Learning for Image Annotation
Chapter 3 Proposed Method
3.1 Graph Regularized Low-Rank Feature Mapping
3.1.1 The Regularization Framework
3.1.2 The Optimization
3.2 Semi-Supervised Dual Low-Rank Feature Mapping
3.2.1 The Regularization
3.2.2 The Optimization
3.3 Robust Semi-Supervised Multi-Label Learning by Triple Low-Rank Regularization
3.3.1 Problem Formulation
3.3.2 The Regularization Framework
3.3.3 The Optimization
3.3.4 APG Algorithm
Chapter 4 Experiments and Analysis
4.1 Image Datasets
4.2 The Preprocessed
4.3 Normalization
4.4 Evaluation Measure
4.5 Experiment Ⅰ: comparisons with the state-of-the-art multi-label learning method
4.6 Experiment Ⅱ: multi-label learning with incomplete training labels
4.7 Experiment Ⅲ: parameter sensitivity
4.8 Experiment Ⅳ: comparisons with the state-of-the-art image annotation methods
4.9 Performance comparison
Chapter 5 Conclusion
References
作者簡(jiǎn)歷及攻讀碩士/博士學(xué)位期間取得的研究成果
學(xué)位論文數(shù)據(jù)集
本文編號(hào):3163928
【文章來(lái)源】:北京交通大學(xué)北京市 211工程院校 教育部直屬院校
【文章頁(yè)數(shù)】:68 頁(yè)
【學(xué)位級(jí)別】:碩士
【文章目錄】:
致謝
摘要
Abstract
Chapter 1 Introduction
1.1 Background
1.2 Problem of statement
1.3 Scope and objectives
1.4 Methodology
1.5 Thesis outline
Chapter 2 Literature Review
2.1 Background
2.2 Automatic Image annotation
2.3 Multi-Label Learning
2.4 Semi-Supervised Multi-Label Learning for Image Annotation
Chapter 3 Proposed Method
3.1 Graph Regularized Low-Rank Feature Mapping
3.1.1 The Regularization Framework
3.1.2 The Optimization
3.2 Semi-Supervised Dual Low-Rank Feature Mapping
3.2.1 The Regularization
3.2.2 The Optimization
3.3 Robust Semi-Supervised Multi-Label Learning by Triple Low-Rank Regularization
3.3.1 Problem Formulation
3.3.2 The Regularization Framework
3.3.3 The Optimization
3.3.4 APG Algorithm
Chapter 4 Experiments and Analysis
4.1 Image Datasets
4.2 The Preprocessed
4.3 Normalization
4.4 Evaluation Measure
4.5 Experiment Ⅰ: comparisons with the state-of-the-art multi-label learning method
4.6 Experiment Ⅱ: multi-label learning with incomplete training labels
4.7 Experiment Ⅲ: parameter sensitivity
4.8 Experiment Ⅳ: comparisons with the state-of-the-art image annotation methods
4.9 Performance comparison
Chapter 5 Conclusion
References
作者簡(jiǎn)歷及攻讀碩士/博士學(xué)位期間取得的研究成果
學(xué)位論文數(shù)據(jù)集
本文編號(hào):3163928
本文鏈接:http://www.sikaile.net/kejilunwen/zidonghuakongzhilunwen/3163928.html
最近更新
教材專(zhuān)著