【作者】 王亮申；

【导师】 欧宗瑛；

【作者基本信息】 大连理工大学 ， 机械设计及理论， 2002， 博士

【摘要】 图像数据库检索是当今信息时代人们广泛关注的热点问题，主要包括对图像内容的描述(特征表达及提取)、图像数据库管理、图像匹配等内容。本文以图像数据库检索为主线，讨论了基于视觉内容的图像检索方法，提出包括基于区域颜色直方图、灰度—基元共生矩阵及向心矩比、偏心矩比、惯性矩比的特征描述方式；对遗传算法存在的早熟、收敛到最优解慢等问题提出了解决方法，并将改进遗传算法应用到图像分割中，编制了相应程序。 颜色特征是描述图像的最常用方法，颜色直方图利用图像颜色的比例分布能够较好地反映图像颜色特征。但是因为颜色直方图很难体现图像的空间信息，会出现同一直方图对应多幅图像问题。本文提出的基于区域块的颜色直方图技术，首先将目标集中在图像的中间区域，再将图像以金字塔结构形式进行分块。通过比较各个区域颜色直方图的相似性匹配图像，可克服全局直方图的局限性，对颜色直方图的度量是其中一个关键技术，利用惯性比来从颜色直方图提取特征，能够保证简单明了地表达颜色直方图。 对富含纹理图像，采用纹理特征表达会更有效，颜色的二维统计方式—灰度共生矩阵是常用方法。对于结构明显的纹理图像可以采用结构方法描述图像特征。本文将统计方法与结构方法有机地结合起来，构建了纹理基元，提出了描述纹理图像的灰度—基元共生矩阵，从这个矩阵中提取用以描述图像的特征值，并组成检索图像的特征向量。 图像分割是解决目标检测、特征提取和目标识别等问题的关键所在，形状特征提取的首要任务就是获得目标图像边界。利用遗传算法对图像能够实现单阈值和多阈值分割。本文首先对传统遗传算法进行了改进，提出了适应函数标定公式，定义了相似度概念。这样可以增加群体多样性，在一定程度上避免早熟现象发生；然后利用改进遗传算法对图像进行了分割计算，将目标与背景分割开来。 形状特征是对图像中边界清晰的目标的最好表达方式。但是即使是找到了目标轮廓，如何来有效地表示该轮廓也是非常困难的。本文提出向心矩比、偏心矩比、惯性矩比概念，对线性化后的轮廓应用本文方法，既能较好地描述图像目标，又具有旋转、平移和尺度不变性。 支持向量机(SVM)是近年来出现的分类方法。本文将SVM及其改进算法SMO应用到图像检索中，可以完成整类检索任务。因为SVM只能够用于两类分类问题，本文构建了二叉树分类器，将SVM技术扩展到多类别分类领域当中；利用前述的图像颜色特征、纹理特征和形状特征作为输入向量，通过训练得到支持向量，构建多类分类函数，进行图像检索。 图像数据库检索系统应能够完成检索、图像数据管理等任务。本文利用VB6.0编制了检索原型系统，包括有特征提取、数据库管理、匹配等模块，采用了基于文本、颜色、纹理、形状和SVM分类检索方法。实验结果表明，本文提出的图像颜色特征、纹理特征和形状特征能够满足对图像内容的描述要求；将支持向量机分类技术应用到图像数据库检索能够得到较好的检索效果。更多还原

【Abstract】 Image database retrieval is a hot topic and has attracted increased attention from researchers. It includes several contents such as describing the image visual content (extracting features), image database management, image matching and so on. This dissertation deals with the content-based image retrieval (CBIR) theory and technique; some new features and tools for more concisely and discriminatingly charactering the content of an image are proposed, such as region-based color histogram, grey-primitive co-occurrence matrix, ratio of centripetal moment, ratio of eccentric moment and ratio of inertial moment. A new modified genetic algorithm is also described in this dissertation, which can upgrade the performance of standard genetic algorithm (SGA) while used in image segmentation.Color feature is often used to describe image content. Color histogram built from cumulative distributions of content colors is a main color feature using in image retrieval. However, the color histogram does not contain the information of spatial distribution of colors across an image. So there might be different images which possess different contents but share with same color histogram. To deal with this problem effectively, a region-based color histogram (RBCH) approach is proposed in this dissertation. In RBCH, an image is first decomposed into several subregions(blocks) with pyramid data structure, and the subregion color histograms are built correspondingly. After that, an integrated color histogram is built based on the subregion color histograms. This integrated color histogram overcomes the disadvantage of simple global color histogram. To digitally represent a color histogram an inertial moment ratio calculated from a color histogram graph is proposed as a new digital feature of the color histogram.It is more efficient to use texture feature to describe an image with complicated textural contents. Grey co-occurrence matrix, which is a 2D color statistical feature, is often used to analyze such images. For the vivid textural structure images it had better use the structural approach to characterize image content. In this dissertation a new statistical tool called as grey-primitive co-occurrence matrix is proposed, which is based on predefined specific primitive texture structures. The features extracts from a set of this type matrix can be formed as a primitive texture feature vector and used in retrieving images from image database. The grey-primitive co-occurrence matrix based approach provides better performance than conventional grey co-occurrence matrix based approach.Image segmentation is very important in object detection, feature extraction and object recognition processing. Object outline detection must be done before shape features are extracted. Recently single-threshold or multi-threshold is often used to segment image and detect object contour on an image by means of genetic algorithm. A modified genetic algorithm is proposed. In the dissertation a different fitness function isconstructed and similarity is introduced, which can increase variety of population md avoid prematurity.Shape feature is the most suitable tool in characterizing image content object which has clear object outline. Even an object contour has been extracted, it is still difficult in constructing suitable shape features to describe the shape discriminatingly and efficiently. Some new features such as ratio of centripetal moment, ratio of eccentric moment and ratio of inertial moment are introduced in this dissertation. These features have the properties of scale-invariance-, rotation-invariance and translation-invariance, and can be used well in depicting the linearized outline object.Support Vector Machine (SVM) is a new classification method. SVM and its implementation technique SMO can be used for image retrieval; however, they were originally developed for dealing with two-classification problem, so a binary tree and a matrix for classification are constructed for solving multi-classification problem. Feature vectors更多还原