【作者】 朱蓉；

【导师】 姚敏；

【作者基本信息】 浙江大学 ， 计算机科学与技术， 2011， 博士

【摘要】 随着数码成像产品和大容量存储设备价格的日益低廉,每天都会产生成千上万的数字照片,导致互联网上多媒体信息的数量呈爆炸性增长。此外,图像搜索引擎(如Yahoo!、Google)和在线照片管理与共享应用软件(如Flickr),促使图像信息资源在网络环境中随处可见。网络用户可以方便地建立自己的数字照片集,并在线与他人共享和交流。但如何高效地表示、索引和检索Web图像,有必要分析图像的内容,挖掘隐含在图像中的语义信息。基于语义的图像分类是从海量的Web图像中发现有用信息的重要途径,具有巨大的应用前景。当前,图像分类虽然在低层视觉特征的提取上较为成熟,但高层语义信息获取的有效性却常常无法满足实际需求。其次,Web图像通常具有大规模、高维数、非线性的特性,并且在内容上又表现出丰富性和多样性。因此,基于语义的Web图像分类既是一个迫切需要解决的关键问题,也是一个极具挑战性的研究课题。本文针对当前Web图像分类中存在的问题,提出了相应的解决方案,获取了一定的创新性成果。本文的主要贡献概括如下：(1)为了缩短存在于图像的低层视觉特征与高层语义信息之间的“语义鸿沟”,首先按照语义信息的抽象程度给出了一个三层的图像语义模型。然后,基于该模型,阐述了用来表述图像内容的具有代表性的语义信息表示方法。最后,从分析Web图像的特点入手,深入研究图像差异性产生的原因,这是获取高效的Web图像分类方法的前提。(2)图像特征优化是处理Web图像分类中的高维图像数据的重要手段。图像特征优化可被形式化定义成一个五元组模型。采用邻域重建和半径递增搜索策略,提出了一种基于球状邻域的局部线性嵌入(GNLLE)。GNLLE不仅适用于稀疏数据集,而且具有较强的抗外界噪声能力和良好的拓扑结构稳定性。由于非线性维数约简通常采用基于点对的相似性计算去定义距离度量,又提出了一种基于球状邻域和路径聚类的局部线性嵌入(GNPCLLE)。由于GPNCLLE充分利用了数据相关性,有效减轻了数据集的整体拓扑结构的失真现象。在Web图像集上的实验结果显示了GNLLE和GNPCLLE的可行性和有效性。(3)准确地从Web图像中提取感兴趣区域(ROI)是提升Web图像分类性能的关键。提出了一种新的图像分割策略,包含粗分割和精分割两个阶段。在第一阶段中,采用基于颜色和纹理特征的块聚类方法将一幅图像划分成四个区域,并依照摄影构图法则将ROI从背景中区分出来。此阶段用来确定目标区域。在第二阶段中,利用形状信息和矢量方法构建了一个活动轮廓模型,图像能量来自于色度梯度,外部能量来自于三角形内心引力和补力。此阶段用来精确地提取目标的边界。通过对Web图像进行实验,验证了提出的图像分割方法的有效性。(4)面向图像集的批处理分类越来越受到研究者的关注。为了提高Web图像分类的准确率,提出了一个建立在图像语义的不同粒度层次上的分层次图像分类模型。定义了两种新的非线性流形：多类对象流形和单类场景模型。在对象层上的分类中,根据不同语义类别图像之间的类内差异和类间差异,利用扩展的局部线性嵌入(ELLE)算法构建了一个多类对象流形。在场景层上的分类中,依照不同场景为同一种语义类别的图像构建了一个单类场景流形,并基于区域增长和线性扰动,提出了一种线性子流形自动生成(LLSE)算法。提出的面向对象的分层次图像分类模型的性能在Web图像集上进行了测试。(5)针对Web图像分类的复杂性,提出了一种基于双流形学习的图像分类方法,将图像分类问题从高维空间转换到低维空间中解决。首先,利用正例类图像和反例类图像之间的显著差异,分开建立两个带有不同固有维数的非线性流形。然后,借助于GNLLE具有的相似性数据聚集的特性,确定各个流形的聚集中心。最后,应用在基于双流形的距离度量上构造的分类器,实现了基于全局的图像分类,克服了建立在整体流形上的分类方法无法反映数据之间真实关系的缺陷。在Web图像集上的实验结果表明了提出的基于全局的图像分类方法是高效的。更多还原

【Abstract】 Nowadays, with digital cameras and mass storage devices becoming increasingly affordable, each day thousands of pictures are taken and multimedia archives on the Internet are growing at an astonishing rate. Besides, image search engines such as Yahoo! and Google, and photo management tools such as Flickr, making image resource available on the Internet. People can easily build personal digital picture collections and online photo sharing has become a common practice. However, for representing, indexing and retrieving pictures on the Internet efficiently, it is necessary and essential to analyze the image content and reveal the semantic information of the images. Semantic-based image classification is an effective way to discover valuable information from a huge number of Web images, having great prospect of applications. At present, the technique of extracting low-level image feature is relative maturity, but the effectiveness of obtaining high-level semantic imformation fails to satisfy the demands of practical applications. Moreover, Web images usually have the characteristics of vast quantity, high dimensionality, and nonlinear structure, and they are also rich in contents. Therefore, semantic-based Web image classification is not only an urgent problem, but also a challenging and rewarding project.This dissertation is aimed to provide a solution to the problem of semantic-based Web image classification. We also have achieved a certain innovative results. The main contributions are summarized as follows.(1) In order to shorten the "semantic gap" between low-level image feature and high-level semantic information, firstly, according to the different abstract degrees of semantic information, we set up a three-layer image semantic model. And then, some representative expression approaches of semantic information for describing image contents are dicussed, based on the proposed model. Finally, with a view of analyzing the characteristics of Web images, we make an intensive study of the causes of the differences between different Web images, which is a premise of developing an efficient strategy for Web image classification. (2) Image feature optimization is an important means to deal with high-dimensional image in Web image classification. Image feature optimization can be formulated via a five-tuple model. We propose a globular neighborhood based locally linear embedding (GNLLE) by using neighborhood update and an incremental neighbor search, which not only can handle sparse datasets but also has strong anti-noise capability and good topological stability. Given that the distance measure adopted in nonlinear dimensionality reduction is usually based on pairwise similarity calculation, we also present a globular neighborhood and path clustering based locally linear embedding (GNPCLLE). Due to its full consideration of the correlations between data, GNPCLLE can eliminate the distortion of the overall topological structure within the dataset. Experimental results on two Web image sets show the feasibility and effectiveness of both GNLLE and GNPCLLE.(3) Segmenting region of interest (ROI) from Web images accurately is important for improving the performance of Web image classification. A novel image segmentation strategy is proposed, which consists of a rough segmentation stage and a fine segmentation stage. In the first stage, an image is partitioned into four regions by using a block clustering based on color and texture features, and the ROI within the image is distinguished from the background according to the principles of photographic composition. This stage aims to determine the approximate region of the target. In the second stage, a novel active contour model is established based on shape information and vector method, where the image energy is defined by a hue gradient and the external energy is generated from both triangular inner force and supplementary force. This stage tries to extract the boundary of the target accurately. Experiments are conducted on Web images to validate the effectiveness of the proposed image segmentation method.(4) The batch-based processing for classifying a set of image sets has been paid more and more attentions. In order to improve the classification accuracy of Web images, we propose a novel hierarchical model based on different granularity levels of image smenatic. In which two kinds of nonlinear manifolds:multi-category object manifold and one-category scene manifold are formulated. The multi-category object manifold is constructed in object-level classification by an extended locally linear embedding (ELLE) algorithm, taking into account both the intra-category and the inter-category discriminations among the images belonging to different semantic categories. During scene-level classification, the one-category scene manifold is set up for the images corresponding to the same object category but having several different scenes. A locally linear sub-manifold establishment (LLSE) algorithm is developed via region growing and linear perturbation. The performance of the proposed object-oriented hierarchical classification model is tested on two Web image sets.(5) As for the complexity of Web image classification, a novel classification method based on double manifold learning is proposed, transforming image classification in high-dimensional space to a lower dimensional feature space. At first, two low-dimensional nonlinear manifolds with different intrinsic dimensionalities are established respectively, according to the significant differences between the positive images and the negative images. And then, each aggregation center of the manifolds is calculated on the basis of the grouping characteristics in GLLE. At last, a classifier is constructed on the distance measure based on double manifold, and it is applied in global-based image classification, having an advantage of reflecting the data topological structure correctly. Experiments on Web images indicate that the proposed image classification method is efficient.更多还原