基于多标记学习的图像标注关键技术研究

发布时间：2018-03-31 10:10

本文选题：图像标注　切入点：多标记学习　出处：《山东师范大学》2016年博士论文

【摘要】：计算机技术和移动拍照技术快速发展,网络空间中的图像信息爆炸式增长。为满足人们对图像的检索,研究人员提出了大量的图像检索算法。图像检索方法可以分为三类,分别是基于文本的图像检索、基于内容的图像检索和基于语义的图像检索。其中,基于语义的图像检索系统中的核心技术是图像的语义标注。本文的重点研究了图像标注的技术问题。研究人员已经提出了大量的图像标注算法,但语义鸿沟问题、维数灾难问题、数据不平衡问题等重要的问题仍然没有从根本上得到解决。针对上述问题,本文基于多标记学习框架,改进了四种经典的机器学习方法用于图像标注,取得了很好的实验效果:1.基于懒惰学习的多标记图像标注算法ML-KNN在计算贝叶斯最大化后验概率时,只使用了主样例与近邻样例在数量上的相关性,没有考虑主样例与近邻样例在距离上的相关性。本文把上述两种相关性同时考虑,提出了一种改进的基于懒惰学习的多标记图像标注算法ML-WKNN。在Image和Yeast两个经典多标记数据集上的实验结果表明,ML-WKNN算法比其它四个经典的多标记算法的总体标注效果更好。2.在基于朴素贝叶斯理论的多标记朴素贝叶斯算法MLNB中,使用主成分分析方法预处理样本的属性特征。处理之后的样例属性之间是不相关的,但是仍然不能满足朴素贝叶斯算法需要属性特征相互独立的要求。本文中我们使用独立成分分析方法来预处理样例的属性特征,处理之后的样例属性特征之间是相互独立的,符合朴素贝叶斯算法对于样例属性特征的要求。在Image和Yeast两个经典多标记数据集上的实验结果表明,IMLNB算法的在多个评价指标上的综合标注效果比其它四个经典多标记算法更好。3.基于改进构建类属属性的思想,本文提出了一种改进的多标记图像标注算法LTFML。LTFML只使用每个类标记的正样例为每个类标记构建类属属性,并使用一种新的评价函数对不同类属属性聚类簇的进行加权。在Image和Yeast两个经典多标记数据集上的实验结果表明,LTFML算法的标注效果在五个评价指标上整体最优。4.针对多标记图像标注中常见的数据不平衡问题,本文对Bagging算法进行改进,提出多标记图像标注集成学习方法BM3。该算法使用Bagging方法对每个类标记的正负样例分别抽取相等数量的样例,然后组成规模相对较小且正负样例完全平衡的训练子集。对基分类器的预测结果集成时,本文使用了一种新的融合策略—最小最大模块化方法。在Image和Yeast两个经典多标记数据集上的实验结果表明,3BM算法整体标注结果比BR等经典的多标记算法的结果更好。
[Abstract]:With the rapid development of computer technology and mobile photography technology and the explosive growth of image information in cyberspace, a large number of image retrieval algorithms have been proposed by researchers to meet the needs of image retrieval. Image retrieval methods can be divided into three categories. They are text-based image retrieval, content-based image retrieval and semantic-based image retrieval. Semantic tagging is the key technology in semantic-based image retrieval system. This paper focuses on the technical problems of image tagging. Researchers have proposed a large number of image tagging algorithms, but the semantic gap problem. Some important problems, such as dimensionality disaster and data imbalance, have not been solved fundamentally. In view of the above problems, this paper improves four classical machine learning methods for image tagging based on multi-label learning framework. Good experimental results are obtained: 1. ML-KNN, a lazy learning based multi-label image tagging algorithm, only uses the quantitative correlation between the main sample and the nearest neighbor sample when calculating Bayesian maximization posteriori probability. The distance correlation between the main sample and the nearest neighbor sample is not considered. In this paper, the above two correlations are considered at the same time. An improved multi-label image tagging algorithm ML-WKNN based on lazy learning is proposed. The experimental results on two classical multi-label datasets Image and Yeast show that the ML-WKNN algorithm is more effective than the other four classical multi-label algorithms. Better .2. in MLNB, a multi-label naive Bayesian algorithm based on naive Bayesian theory, The principal component analysis (PCA) method is used to preprocess the attribute characteristics of the sample. In this paper, we use independent component analysis (ICA) method to preprocess the attribute features of the sample, which is independent of each other. The experimental results on two classical multi-label datasets of Image and Yeast show that the algorithm has more comprehensive labeling effect on multiple evaluation indexes than the other four classical ones. The tagging algorithm is better. 3. Based on the idea of improving the construction of generic attributes, In this paper, an improved multi-label image tagging algorithm, LTFML.LTFML, is proposed to construct class attributes for each class tag using only positive samples of each class tag. Using a new evaluation function, the clustering of different generic attributes is weighted. The experimental results on two classical multi-label data sets, Image and Yeast, show that the tagging effect of the algorithm is the best in five evaluation indexes. Aiming at the problem of data imbalance in multi-label image annotation, In this paper, the Bagging algorithm is improved, and an integrated learning method BM3 is proposed. The algorithm uses the Bagging method to extract an equal number of positive and negative samples from each class label. Then the training subset with relatively small scale and complete balance of positive and negative samples is formed. When the prediction results of the base classifier are integrated, In this paper, a new fusion strategy, minimum maximum modularization method, is used. The experimental results on two classical multi-label datasets, Image and Yeast, show that the global labeling result of the algorithm is better than that of classical multi-label algorithms such as Br.
【学位授予单位】：山东师范大学
【学位级别】：博士
【学位授予年份】：2016
【分类号】：TP391.41

【相似文献】