【作者】 于永新；

【导师】 冯志勇；

【作者基本信息】 天津大学 ， 计算机应用技术， 2009， 博士

【摘要】 图像的语义识别和检索,一直以来都是计算机领域的热点问题。该问题涉及了图像处理、模式识别、人工智能以及机器视觉等众多学科领域。本文针对其中的若干关键问题做了研究,取得了如下一些成果:提出了图像语义的四层模型和认知架构。在综述当前研究现状的基础上,提出了由低到高分别为:特征层、实体层、关系层、语义层的四层模型。在此基础之上,我们提出一个基于本体的图像语义认知架构。该架构共由五个模块构成:图像分割模块、实体库构建模块、决策树归纳学习模块、实体解析模块和高层语义解析模块。该系统架构能够对于图像进行区域分割,并依据训练的判定算法,映射得到区域的实体概念。可以结合OMCSNet常识库,进行语义消歧。构建了图像实体模板库。采用Berkeley人工分割图像的思想,在ImageJ框架下,以JAVA插件的方式实现了低层特征的获取。对于模板图像,分别提取颜色特征值(Lab颜色均值)、纹理特征值(灰度共生矩阵的熵、对比度、能量和反差分矩)和形状特征值(Hu形状不变矩)。提出了基于各向异性扩散滤波的区域增长分割方法。在各向异性扩散滤波中,使扩散控制参数自适应于图像的粗糙度,以达到对高粗糙度的纹理区域,具有较强的扩散作用。经过各向异性扩散滤波,使得图像整体比较平滑的同时,却能保持图像的边缘比较清晰。经过滤波后,使用了区域增长的图像分割方法。并使用经过归一化处理的混和距离(颜色距离和粗糙度距离相组合)来对区域进行合并。并使用决策树归纳算法,构建了实体的判定规则。定义了图像本体,提出了基于OMCSNet语义网的语义消歧方法。利用低层特征描述表,定义了图像本体。同时,本体中描述了实体之间的空间拓扑关系,并给出了基于SWRL规则语言的空间关系定义。为了达到在上下文环境下的语义消歧,使用了OMCSNet语义网,并给出了基于OMCSNet的语义消歧原理。更多还原

【Abstract】 The semantic recognizing and retrieval of image has being a focus in computer academic field, which includes image processing, pattern recognizing, artificial intelligence, machine vision and so on. Some key problems are studied in this thesis, and some results are obtained.A four-layers image semantics model and recognizing framework are put forth. After a review on image semantics and recognizing, a four-layers model is introduced, which is composed of four layers from bottom to top: feature layer, entity layer, relationship layer, semantics layer. Then a semantics recognizing framework based on semantic network is proposed, which is composed of five parts: image segmentation, entity lib building, decision tree study, entity parsing, semantics parsing. The framework can do image segmentation and map region to entity based on decision rules. Semantic clarifying is also can be done by OMCSNet.Image entity template lib is built. According to Berkeley artificial segmentation lib, a java plugin is designed to obtain low level features based on ImageJ framework. For template image, the following features are produced: Lab color mean value; the entropy, contrast, energy and IDM of gray level concurrence matrix; Hu shape invariable moment.A region growing segmentation method is introduced which is based on adaptive anisotropic diffusing filtering. A diffusing control parameter is built to adaptive to the coarseness of image, which can make stronger diffusing effect in high-texture region. After adaptive anisotropic diffusing filtering, image become smooth, while the edge of region remain sharp. Then a region growing method is used to segment the image, which uses the combing distance of color distance and coarseness distance to merge small regions. Decision tree is built to generate decision rules.Image ontology is built, and semantics clarifying method is put forth based on OMCSNet. An ontology of image is built based on low level features description table. Space topological relationship of image ontology is defined by SWRL rules. To obtain the true meaning of region in image, OMCSNet is adopted to clarify the meaning of a region.更多还原