【作者】 谭立球；

【导师】 谷士文； 夏利民；

【作者基本信息】 中南大学 ， 控制理论与控制工程， 2009， 博士

【摘要】 随着计算机技术和国际互联网的飞速发展,包括图像在内的各种多媒体数据的数量正以惊人的速度增长。如何有效、快速地从大规模图像数据库中检索出所需要的图像是目前一个急需解决的重要问题。人们对于图像的理解,往往是建立在基于视觉的基础上,基于内容的图像检索存在低层语义与高层语义的语义鸿沟,而基于语义的图像检索是以单个概念为基础,并没有考虑概念之间的彼此关系。基于此,本文将本体论引入到图像检索中,提出了基于本体的图像检索方法,主要工作如下:1.在综述了基于内容、基于语义的图像检索相关技术的基础上,将领域本体引入图像检索中,提出了基于本体的图像检索框架,在该框架下图像描述包含了视觉低层特征、高层语义概念,既能充分利用图像本身的低层特征,又能符合人的图像视觉理解;通过本体,可以定义图像的语义概念之间的关系;同时通过本体还可以进行语义扩展,弥补语义查询过程中的信息不足。2.将信息瓶颈算法用于图像分割,提出了基于信息瓶颈算法的图像基元提取方法,考虑到图像中距离接近的两个区域很有可能属于同一个图像基元,而远离的两个区域则很可能属于不同的基元,在使用凝聚的信息瓶颈算法对图像像素进行聚类时,同时考虑互信息的损失和聚类区域之间的空间距离,以得到更有效的图像基元。该方法与传统的聚类方法相比,其聚类的结果与距离函数无关,且不依赖于初始聚类中心的选取。实验验证了算法的可行性和有效性。3.提出基于本体的图像自动标注方法,在训练阶段,采用基于语义约束的半监督信息瓶颈聚类方法对提取基元进行聚类,对信息瓶颈聚类算法进行了改进,提出了基于半监督约束聚类的信息瓶颈算法,使用少量的标记样本来帮助无监督的学习,将特定的一已知知识以“约束”的形式表达,并嵌入到聚类过程中的方法,使得聚类算法获得更多的启发式信息,提高了效率和聚类质量。采用统计法和半监督学习方法建立了图像基元类与本体中语义概念关系概率表。在自动标注阶段,采用二次标注方法实现对于图像语义的自动标注,首先通过分类方法获得获得图像属性概念标注,然后结合本体知识,获得图像的概念标注。实验验证了方法的可行性和有效性。4.定义了图像本体框架下图像相似度模型,并给出了基于近似向量的相似度计算方法;提出了在本体框架下基于LPP VA-File的图像快速检索方法,在保局投影变换域中建立近似向量文件,通过保局投影消除了原图像特征数据各分量之间的相关性,并保留了图像数据的非线性特性;给出了图像检索K近邻查询算法,该算法减少了对原图像特征向量的访问数量,即降低对原图像特征向量的I/O访问时间,大大提高搜索效率。实验验证了方法的有效性。5.提出了基于结合先验知识SVM的相关反馈方法。结合领域本体的先验知识定义了新的带权值的训练样本,解决了支持向量机训练过程中已知样本少的问题,样本的权值反映了本体的先验知识和用户的关注兴趣,权值越大的样本说明该样本可信度越大,在支持向量机学习中所起作用的越大;提出了结合先验知识的支持向量机;在此基础上提出了基于结合先验知识的支持向量机的短期学习和长期学习的相关反馈算法。更多还原

【Abstract】 The progress of computer technology, multi-media technology and internet technology result in the explosive growth of image.How to fully make use of the image and get useful image is the hot research. Traditional image retrieval can not meet the needs understanding in the semantic level. We focus on ontology based image retrieval and the main contributions are following:1 .After surveying the research state of the content based image retrieval and semantic based image retrieval, an ontology based image retrieval framework is proposed which domain ontology is used. This framework covers the visual feature and semantic concepts. it can make fully use of image visual feature and conform to human visual understanding. Ontology not only can define the concept relation ,but also can make the query information insufficiency during the image retrieval.2.A method of information bottleneck is proposed to image segmentation, and it extract the image blobs using information bottleneck .Considering the two regions whose are the neighbour may in the same blobs , and the remote ones may belong to two difference image blobs ,an agglomerative information bottleneck method is applied to cluster image pixel which takes mutual information cost and the cluster region space distance for getting the better result. Comparing with other cluster methods, the cluster result doesn’t rely on distance function and initialization cluster center. The experiments validate the feasibility and validity of the method proposed.3.An ontology based image automatic annotation algorithm is presenet. In training phase ,adopt semi-supervised information bottleneck algorithm to cluster the blobs .It use small marked samples to help unsupervised learning. During the clustering process ,it use some priori knowledge as constraints ,and can get more heuristic information, and improve the efficiency and cluster quality. A probability table of the blobs and ontology concept is constructed using statistics and semi-supervised learning.In automatic annotation phase , use two steps to annotate images, first get the image attribute concept using classification , then get the image concept from ontology.The experiments validate the feasibility and validity of the method proposed. 4.The image similarity model is defined using image ontology framework, and the similarity algorithm based Approximation vector is given; A rapid retrieval method based LPP VA-File is proposed under ontology framework, which construct VA-File in Locality Preserving Projections transform domain, eliminate the relativity of image feature vector, and keep the nonlinearity; KNN under LPP VA-File is present which can eliminate the accessing of original image feature , reduce the I/O time with orginal image, and retrieval efficiency is improved greatly. The experiments validate the feasibility and validity of the method proposed.5.A new relevance feedback method using SVM with priori knowledge is proposed in this paper. The training samples is defined with new weights using ontology priori knowledge ,which overcome the the deficiency of small samples in SVM training. The weight of samples reflect the ontology priori knowledge and users’ interests. The higher the weight ,the more reliability the samples will have ,and they will have more effect in SVM. SVM with priori knowledge is proposed ,and the short relevance feedback and on which the long relevance feedback are proposed.更多还原