【作者】 石跃祥；

【导师】 蔡自兴；

【作者基本信息】 中南大学 ， 计算机应用技术， 2005， 博士

【摘要】 计算机视觉是模式识别应用的一个重要方面,计算机视觉的目的是通过电子化地感知和理解图像,复制人类的视觉效果。计算机视觉首选要解决好三维物体所形成的二维图像在计算机中的存储、处理、描述、表达和理解操作等问题,它涉及的知识广、内容多,包括光学、物理学、数学、自动化和计算机等学科。计算机视觉就是在多学科的基础上来综合各种知识与技术方法,形成一个具有视觉功能的信息处理系统,从而建立起计算机对外界世界的感观、认知和行为动作,即计算机视觉。 本文针对计算机视觉的图像语义模型方法进行了研究。通过分析和研究目前国内外计算机视觉的现状、主要问题和未来发展的方向,对于计算机视觉中图像识别知识表达的问题,特别是图像识别中的底层属性与高层信息关联等方面的内容,进行了设计与研究,构建了计算机视觉图像语义模型方法的知识结构,为图像识别建立了知识关联,实现计算机视觉图像识别的知识方法。 针对计算机视觉图像语义的问题,本文提出了基于知识结构与方法的图像语义模型理论和方法。论述了语义名词的内涵和概念表达,阐述了图像高级知识相关属性与语义的关系、连接和范畴,定义了图像语义名词概念,建立了图像语义模型的理论基础。 提出了图像语义的特征空间表示法。它建立了图像属性与信息描述的映射关系,描述了图像底层属性、图像对象关系及对象空间关系和图像高级信息的信息映射过程,从三层结构的映射来形成一个基于知识描述图像内容的特征向量。图像语义反映了图像存在的基本特性,即图像的光表现,图像中对象的关系,即空间关系,图像的高级信息,即图像与社会与人的关系。从高级信息的知识角度来分析图像存在与应用的环境问题,解决了图像应用信息在计算机中描述与表达的相关知识问题,为计算机视觉的图像识别应用,建立了一种新的知识方法,提供了可参考的思维方式和结论。 计算机视觉图像语义模型描述了在设定的场景中,对在计算机视觉中直接成像的目标物体图像的描述,对所获取的图像,选择相应属性和对象关系等信息,利用约束机制,将数据信息映射到特征空间上,从而形成一个对图像的描述模型。为了建立图像属性与高层语义之间的关联,提出了基于SVM的图像语义关联法,它描述了图像底层属性到高级描述之间的关联,实现了图像信息的映射和特征表达,解决了图像属性选择和高级描述之间的关联问题。 在计算机视觉系统图像语义模型应用中,解决了语义概念运用和语义特征值表示、数字化的问题,建立了图像语义的数据模式和图像语义数据库。提供了图更多还原

【Abstract】 The research on the desceiption of image semantic models in computer vision is a very important work. When the computer assembled with vision function, it can help people in all fields. This vision aims to the solution of storage, process, description, expression and understanding image in 2D to 3D objects. The vision system includes based knowledge on Optics, Physics, Mathematics, Automatics and Computer Science. The research integrates each course to generate a new branch for vision. The computer will get the ability of sense, recognition and decision from the integrated.The paper presents a concept and application in computer vision with high knowledge. It has analysed the existing state of computer vision, causes and the demands of development in the future. Especially in the image processing of computer vision, the paper organized the basic knowledge of methods and principles on image processing. It also gave the experience on description, content, and the detail in computer vision of image processing. It can impress studying people a resume and outline about computer vision. Helping human to learn the knowledge about computer vision.In order to slove the problems of self-learning and understanding, computer vision should be constructed on high knowledge molds. The paper presents a method which using the knowledge of image semantics and theory. The semantic concept was introduced with connotation and evolution. This discuss includes the relationship between high knowledge and semantics, linking, category. The summery is a base to the word "semantics" which will be frequently used in computer vision.Semantics is a high description of image attributes. The description is a hierachical structure that contains three layers. This construction extracts the information from the image attributes, relation of objects in space and high information of image. A feature vector which be used in the recognition of vision can be extracted from the three layers. This vector was used to describe image in high knowledge. The semantics presents relations and structures of image and spectral, object and object, image and sensibility. The relation of objects reflects location and state. But the impression refers the image in the society. This research is to find the reason on image reserve and condition in high knowledge. Seeking a new field to serve the society using the rich image resource. This research is an advanced and necessity for the image application using a new idea and efficient results.The semantic definition and its concepts have been given in beginning of the paper. The related events with semantics also have been interpreted. Image semantics of computer vision is the means of changing the recognition from known to unknown environment. The description is the object that projects to the computer vision/camera in direction. The data was epurated from selected object or objects in image attributes. There is an eigenvalue that has been extracted by the projection of vector space. This discuss mainly connects with the basic application. The description was formed like the projection from image attributes/bottom to high information. The projection is the solution to how to select the image attributes and construct the semantics. In order to realize this, the SVM has been used to build the relation between the attributes and senior information. The principle and methods of generation also have been constructed followed by these definitions. This is the basic theory system of computer vision.In the paper, it also gives the related solutions that how to use the concepts of semantics and the digited description in the application of computer vision. The eigenvalue was given to represent the semantics to feature vector space. There are necessary to offer the pattern recognition and database. In order to describe the semantics, the symbols strings were used to represent the eigenvalue from semantic models. In the operation, the recognition of models and database have also been defined and settled. This work is useful to solve the presentation of semantics, category of concept, models of concept, process and steps. Data has been managed in database.The paper presents the representation of semantic models and semantics structures to heighten the efficiency and reliability of image semantic application. It also solved the key question of semantics storage in the computer. The presentation showed the related complex and experimental results. The semantic models, structure of semantic models and structure of semantic storage have been constructed in high knowledge. The structure of storage was built in cross link-list. This strategy comes from the application and the concept-self. The description of third sects was presented in eigenvalue and characters. The structure reflects the feature of semantics. The link-list can easy realize the linked operation and semantic operation between different semantics. The cross link-list can present the relation and knowledge expression. This storage structure offers a stage for self-study and reform.The rules of semantic generation were constructed for getting semantics. Basedon the rules, generating a semantic just like to extract the eigenvalue from the feature tree. The semantic tree is called description of model(TDM). The prior of depth and clustering were considered to get the semantics from the TDM has been suggested. The threshold M is also considered to use in the clustering to reduce numbers of element. These made the process of generating semantics clear and sample. This work helps people to analyse and find errors in the semantics. The similar computation has been designed to offer the degree for the object in semantic recognition.The paper presents the knowledge operation of image semantic models to solve the problems which how to show the relation between the semantics. From the construction of semantic knowledge and feature vector, the structure can be defined and to be used in semantic operation. In order to manage and operate the knowledge structure, the semantic models and operations have been classified. The operation BNF structure has also been defined. The related degree calculates the relation among the semantic models.There are two kinds of semantic experiment. This strategy is to divide the solution into two parts. The semantic models were designed to bridge image and high knowledge. The semantic application began with the semantic models. Then the experiment constructed the description in high knowledge based on the semantic models. From the experiment, track of human motion and track of road automobile, the result showed the two kinds would be easy accepted and realized. The results compared with that of general method that did not used the semantic model is wonderful. Based on the above, the high knowledge of semantics was efficient to construct eigenvalue in the emulational experiment. The results present the innovative and advantaged of semantic application.Based on the summary, there are some results and prediction for the semantics study and application in the future.The paper has constructed and added some description for computer vision in knowledge structure. This point is to aim the high knowledge to solve some problems in image retrieval and recognition in the computer vision. This work is significance in theory and application. The study exploits a new description for computer vision.更多还原