【作者】 曾璞；

【导师】 吴玲达；

【作者基本信息】 国防科学技术大学 ， 控制科学与工程， 2009， 博士

【摘要】 随着数字成像技术的快速发展,数字图像的数量也在飞速增长。越来越丰富的图像资源使用户难以在浩如烟海的图像数据中找到真正需要的图像信息,因而,如何实现快捷、高效的图像组织与检索就成为颇具价值的研究课题。近年来,通过图像分类来提取图像语义内容已成为被广泛关注的研究热点问题。然而,目前面向语义提取的图像分类技术面临诸多挑战,如何构建有效的图像分类方法仍是一个值得深入研究的问题。本文围绕语义提取需求背景下的图像分类研究这一主题,主要针对特征提取、多特征融合和多类分类器设计等关键技术展开研究。论文的主要工作与创新体现在以下几个方面:1、提出了一种面向场景分类的图像区域潜在语义分布特征提取方法。该特征提取方法的核心在于区域潜在语义的获取,具体过程是首先采用空间金字塔分块生成图像分块区域,然后对图像分块区域集合应用概率潜在语义分析方法自动挖掘出区域潜在语义,最后联合所有图像分块中区域潜在语义的出现概率来构建区域潜在语义分布特征。与其它中间语义特征相比,该特征在不需要人工标注的情况下能利用图像区域语义的空间分布为场景分类服务。在13类场景图像上的实验验证了该特征的有效性。2、提出了一种基于核函数组合的多特征融合分类模型及其优化方法。该方法通过特征所对应核函数的线性加权组合来实现特征融合,将最优多特征融合问题转化为组合核函数的优化问题,并利用多核学习方法进行优化。为克服传统的多核学习方法只满足类间间隔最大化要求的缺点,根据Fisher准则提出了一种基于类间间隔最大化和类内散度最小化的多核学习算法,并将其用于多特征融合分类模型的优化。实验表明,采用该方法优化的多特征融合分类模型能同时实现特征选择和特征融合,具有更好的分类性能。3、提出了一种级联模型子空间最小距离分类器的多类分类方法。该方法的核心是通过最小距离分类器筛选出较小的候选集以进行最终的分类。为保证该方法的分类速度和分类精度,提出在模型子空间应用最小距离分类器,并且通过构建一个基于权值稀疏性约束的距离度量学习方法来同时获得最优的模型子空间和相应的距离度量。该方法在不损失分类精度的前提下,能大幅提高分类速度,在Caltech256数据集上的实验验证了这一结果。4、提出了一种融合特征分布和类别语义相似度的层次分类器生成算法。基于特征分布的类别相似度是在获取基于聚类的类别概率分布基础上,利用概率分布之间的距离来实现;而基于语义的类别相似性度量则是利用类别词在WordNet上的语义关联度来实现。最终,通过二者的线性融合来生成类别相似度,并且利用谱聚类方法实现层次分类器的构建。该方法能利用类别特征分布信息和语义相关信息之间的互补性为构建层次结构服务。与只使用一类信息的方法相比,具有更好的分类性能。更多还原

【Abstract】 With the rapid development of digital imaging technology, the amount of image is increasing rapidly. It’s difficult to find the user-wanted images from huge mount of image data. So how to organize these images and retrieval a special image from the mass database efficiently and effectively has become a major issue. Image classification is an important and challenging task in this field and is attracting more and more attention. But there is so much difficulty in image classification task for semantic extraction, how to generate a more effective image classification method is still an open problem.In this dissertation, some key techniques of image classification for semantic extraction have been explored, which include feature extraction, multi-feature fusion and multi-class classifier. The original contributions of this thesis can be described as follows:1. An image regional latent semantic distribution feature is proposed for scene classification. The core of this feature extraction method is how to get the regional latent semantic. Firstly, an image block collection is generated by using spatial pyramid subdivision method on training image collection. Then the Probability Latent Semantic Analysis method is used on the image block collection to mine the regional latent semantic. Finally, the image regional latent semantic distribution feature is defined by uniting the probability value of each regional latent semantic in each image block region. Comparing with other intermediate semantic features, this feature has used the distribution of regional semantic to improve image classification performance, as well as it reduces the load of people. Experiment results show that this feature has satisfactory classification performance on a large set of 13 categories of complex scenes.2. A multi-feature fusion model based on kernel combination and its optimization algorithm is proposed. In this classification model, multi-feature fusion is completed by a convex combination of feature kernels, and each feature kernel corresponds to an image feature. Then, the problem of how to fuse image features excellently has become another problem which is how to optimize the combinational kernel. To solve this problem, multiple kernel learning can be used. But the multiple kernel learning methods in existence have only maximized the between-class variance. Based on Fish rule, an excellent classifier should maximize the between-class variance and minimize within-class variance. To satisfy this rule, a new multiple kernel learning method is proposed and has been used to optimize multi-feature fusion model. The experimental results show that the proposed method can finish feature selection and feature fusion at the same time, and has higher classification accuracy. 3. A multi-class method cascading minimum distance classifier in model subspace is proposed. Firstly, this method uses minimum distance classifier to get a litter class collection, and then multi-class SVM classifier is used to classify on this collection. To improve the classification performance, a minimum distance classifier based on model subspace is proposed. To guarantee the classification speed and precision of minimum distance classifier on model subspace, a new distance measure learning method which based on sparse restriction of distance weight is proposed. By this method, an optimization model subspace and an optimization distance can be generated at the same time. The experiments on Caltech256 show that the proposed method can guarantee the classification speed and precision.4. A new inter-class similarity combing feature distribution and class semantic is proposed, and this measure has been used to automatic generation of class hierarchy. In the measure based on feature distribution, the train data has been clustered firstly, and then the prior probability distribution of cluster has been used to describe the feature distribution of each class, finally a distance based probability distribution is used to get the inter-class similarity. On the other side, the semantic similarity of class word has been computed based on WordNet. The total inter-class similarity has been computed by combining these two measures and a class hierarchy has been automatically generated based on this measure by spectral clustering algorithm. Comparing with the method only using one measure, the proposed method has higher classification performance.更多还原