【作者】 张恒博；

【导师】 欧宗瑛；

【作者基本信息】 大连理工大学 ， 机械设计及理论， 2008， 博士

【摘要】 图像数据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、图像数据库管理、图像匹配等内容。本文从图像数据不同特性出发,讨论了基于视觉内容的图像检索方法,提出包括基于主色调多特征向量、综合灰度颜色和空间特征、综合颜色和纹理特征、基于形状特征等多种图像检索算法。随着图像数据库规模的越来越大,如何提高图像检索的效率已成为迫切研究的问题。颜色信息是图象中最为明显有区别性的信息,但采用单一的图像特征向量对图像数据库进行查询不能很好地解决查询中准确率和效率之间的矛盾关系。为此本文提出了一种利用多种特征向量的彩色图像检索方法。该方法基于HSV颜色模型提取示例图像的颜色特征与图像库中图像的颜色特征时,提取不同维数的特征向量。在图像匹配时对不同维数的特征向量采用不同的距离计算方法。并且通过设置阈值实现分层检索图像数据库。从而提高了图像检索的效率。由于颜色直方图只记录了全局的颜色统计信息,未包含颜色的空间分布信息并又混入了不感兴趣物体的颜色信息。针对上述的问题,本文提出一种综合图像的颜色信息、灰度信息和空间信息提取图像的特征向量的方法,首先将图像从RGB空间转化为HSV空间,并进行非均匀量化为32种代表色:其次将彩色转化为灰度图像:然后将图像划分成互有重叠的8块,分别求出每块图像的颜色直方图、灰度直方图;采用这16个直方图做为图像的特征向量对图像进行检索。纹理是图像的一个非常重要的特征,但是单纯依靠纹理特征进行图像检索,其应用范围较窄,还需综合考虑图像的颜色及其分布等其它视觉信息。本文提出一种结合图像颜色空间分布信息及其纹理特征的图像检索新方法。该方法基于图像的颜色连通区域,结合图像的颜色构成和分布信息,提取图像的多个颜色分量的共生基元矩阵纹理特征。然后,利用针对该特征的图像相似性度量函数实现基于内容的图像检索。基于形状的图像检索一直以来是图像内容检索的一个难点问题,本文提出了一种新的针对图像形状的检索方法。首先,用Canny算子对图像进行平滑处理,提取图像边界方向直方图特征。其次,用改进的不变矩来描述图像形状的区域特征,改进后的不变矩特征不受图像的缩放、平移和旋转的影响。最后,为了克服不变矩只关心对象区域,而对图像边界忽视的缺点,提出了改进的不变矩与边界方向特征相结合的方法,使得检索取得更好的效果。更多还原

【Abstract】 Image database retrieval is one of the hot topic and has attracted increased attention from researchers. It includes several contents such as describing the image visual content, image database management, image matching and so on. This dissertation takes the image database retrieval as a master line, discusses the image retrieval method based on the visual content, proposes new image retrieval algorithms including algorithms based on the host tone multi-characteristic vector, the color and the spatial combination characteristic, the color and the texture combination characteristic, the shape characteristic and so on.Along with image database getting larger and larger, how to enhance the image retrieval efficiency has been become an urgent research issue. It is not able to yield a good result by using the sole image characteristic vector as image feature for inquiry. The retrieval proceding should get a satisfied balance between the rate of precision and the recall rate, in view of it, this dissertation proposed a new image retrieval method based on the multi-characteristic vector of color image, the features were extracted and processing on HSV color model and clustered to form different dimension characteristic vectors. A hierachical mapping schema with different dimension characteristic vectors is presented, which does enhance the image retrieval efficiency.To further upgrade the retrieval performance, a new kind of characteristic feature vector which is composed of color, intensity and spatial distribution information, is also proposed in this dissertation, the new features yield better retrieval results.Texture is an extremely important characteristic of images; however, it is not very efficient in pratical applications if only sololy using texture characteristics features. It is expected that the texture feature information is combined with color information, the information spatial distributions of colors and other visual information. This dissertalion proposes a new kind image retrieval method with combination of image color spatial distribution information and the texture characteristic. Firstly, the image was divided as several parts, then the regions connected with similar color were clustered, and the primitive co-occurrence matrix of four colors are extracted, finaly the image retrieval on the basis of the content is realized through using the feature similar criteroon which is designed according to these features.Since image retrieval based on shapes has long been one of the difficult problems in content based retrieval. The dissertation proposed a new kind image retrieval method based on shape content. First, a Canny operator is implemented to smooth the source image and get the feature of the edge direction histogram, then the improved invariant Moments are calculated for describing shape feature. These features are invariant under rotation, scale, translation and reflection of images and have been widely used in a number of application due to their invariance properties. In view of that the invariant moments focus on the distritntion of the region areas and in lower level reflects the shape contour characteristic, we propose a new method based on the integrated informations of invariant moments and edge direction.更多还原