【作者】 李小燕；

【导师】 程显毅；

【作者基本信息】 江苏大学 ， 模式识别与智能系统， 2008， 硕士

【摘要】 目标的表示和识别技术是图像分析和理解的核心环节,合适目标的表示是基于内容的图像检索的基础,不同的表示方法将导致不同的识别方法。目前已经有了不少目标表示方法,但是由于缺少结构化的表示,使得基于内容的图像检索结果不能令人满意。为了解决这个问题,朱松纯和大卫·孟弗德提出了在“与或图”中嵌入随机文法表示对象的图文法思想,并综合了一些如马尔科夫随机场和稀疏编码等这样流行模型产生了一个学习结构实现对文法的推断,但是这种方法缺少图像的基元之间的语义信息。由于目标骨架具有层次化、多尺度、与原始目标拓扑一致和适应较大的变化等特点,因此基于骨架的目标表示和识别开始受到人们的关注。本文结合了随机文法和骨架化的优点提出了用随机文法进行骨架结构化表示,增加了图像基元之间的语义特征,有利于提高基于内容图像检索的精确性。本文的主要工作:(1)针对传统骨架提取算法中出现影响骨架识别的毛刺问题,特别是其中对物体形状的描述会产生很大影响的绷带骨架,在骨架的权值的基础上提出了一种新的骨架修剪算法。(2)结合骨架和随机文法提出了用随机文法进行骨架结构化表示模型SGIRS(stochastic grammar on image representation based on skeleton)。(3)利用了骨架的柔性以及随机文法的抗干扰能力,提出了基于SGIRS的目标识别框架,并在Mpeg-7所给测试集的两个形状图像库上进行了实验,实验结果表明本文的方法比不考虑骨架分支权值的方法在降低了目标主要形状丢失的概率方面有效。更多还原

【Abstract】 Object representation and recognition techniques are kemel issues in the image analysis and understanding,in which the appropriate object representation is the groundwork and different representation forms will result in different recognition strategy.The result of content based image retrieval is dissatisfactory due to the lack of structural representation.To solve this problem,song-chun zhu and David Mumford proposed to express object by embedding a stochastic graph grammar in an And-Or graph,and unified a number of popular models in the literature,such as Markov random fields and sparse coding with wavelets to realize grammatical inference,but method is lack of semantic information between the image.primitives.The representation and recognition approaches based on skelet on are recently paid more attention since skeleton has following characteristics:hierarchical, multi-scale,topology of uniformity and adaptability of variety.This paper combines the advantages of stochastic graph grammar and the skeleton,and proposes ideal using stochastic graph grammar to express skeleton structure,so it will help increase the semantic features between the image primitives and improve the content-based image retrieval accuracy.This paper main work concentrates in:(1)Aimed at the appearance of burr in the skeleton extraction algorithm,especially the ligature which has a significant impact on object shape descriptions,proposing a skeleton pruning algorithm based on the skeleton weight.(2)Combines the advantages of stochastic graph grammar and the skeleton, proposing a model SGIRS(stochastic grammar on image representation based on skeleton).(3)Taken advantage of flexibility of stochastic graph grammar and anti-interference ability of skeleton,proposing a object identification framework based on SGIRS and the experiment on two shape librarys given by Mpeg-7-test,the results show that this method better than not considered the weight of skeleton at reducing the probability of loss main shape for objective.更多还原