基于距离测度学习的图像分类方法研究

【作者】 张杰；

【导师】 翁巨扬；

【作者基本信息】 复旦大学 ， 计算机应用技术， 2010， 硕士

【摘要】 基于内容的图像分类方法是计算机视觉领域中的难点和热点问题之一。目前几乎所有的图像分类方法都依赖于用图像底层特征间的距离来度量图像内容的语义相似度,实现对图像内容的理解。然而,图像的底层特征和高层语义间往往存在巨大的语义鸿沟。如何缩小语义鸿沟,提高图像分类系统的性能,是一个得到广泛研究的课题。此外,由于人们对于图像内容间的相似度往往带有主观性,如何将这些影响反映到图像间特征相似度的计算上,实现“个性的”图像分类是一个值得深入研究的方向。本文首先介绍了图像分类系统的一般框架,图像的视觉底层提取,重点介绍分析了不同特征间的距离度量方法。一些距离度量方法由于缺乏对特征空间较好的描述,忽略了背景环境信息对距离度量的帮助,往往效果不是很理想。因此,如何能够从有类标的图像特征空间中自动学到一种合适的距离测度就显得十分关键。本文对距离测度学习方法做了系统的分析研究,探讨了不同距离测度学习算法的本质以及相互之间的联系。接着,本文提出了一种基于特征互补率矩阵的图像分类方法,试图从两个角度减小语义鸿沟。一方面,图像的各种视觉底层特征描述了图像内容不同方面的特性,因此可以通过融合这些特征更好的描述图像内容,这里我们通过计算特征互补率矩阵来指导融合特征集的选择；另一方面,对线性融合后的特征,利用距离测度学习算法从图像训练集中学到一个距离函数以更恰当的反映图像内容间的相似度。实验表明,该方法能有效提高分类系统性能。此外,我们通过实验研究了图像数据集不同语义粒度的划分对分类性能的影响,比较分析了几种分类算法以及局部特征在图像分类任务中的性能,评估了十种MPEG-7定义的视觉描述子的分类性能。更多还原

【Abstract】 Content-based image classification is one of the hot and knotty issues in computer vision. Recently, almost all current approaches rely on distance between low-level features for judging semantic similarity, and then understand the content of image. There is, however, a deep semantic gap between low-level features and high-level concepts. The topic on how to bridge the semantic gap effectively and improve the accuracy of image classification has been widely explored. Moreover, the subjectivity in similarity needs to be incorporated more rigorously into image similarity measures, to achieve what can be called‘personalized’image classification.In this paper, a general framework of image classification system and extraction of low-level feature, especially distance measures of various features are introduced. Some distance measures have poor performance in the task of classification, since they fail to describe the feature space and neglect the help of contextual information. Hence, the key point is how to learn a proper distance metric automatically from the feature space of labeled images. This paper analyzes the methods on distance metric learning, and explores the nature and connection between them.Then, we propose an image classification method based on feature complement ratio matrix, which tries to bridge the semantic gap from two aspects. On the one hand, different low-level feature describes different characteristic of the content of image, so a better representation can be obtained by fusing these features. A visual feature complement ratio matrix is been computed, based on which the fusion feature set can be selected effectually. On the other hand, a proper distance metric, which reflects the semantic similarity between images, is obtained by using a distance metric learning algorithm. Experiments show that our method can improve the accuracy of image classification effectively.In addition, some experiments have been conducted on analyzing the effect of different division of image set and the performance of several classification methods, local feature and ten MPEG-7 visual descriptor in the task of image classification.更多还原