【作者】 李齐周；

【导师】 陈文兵；

【作者基本信息】 南京信息工程大学 ， 应用数学， 2011， 硕士

【摘要】 基于内容的图像检索(CBIR)是利用图像本身的信息,借助现有的图像处理技术和构造新的算法来辨别图像特征的机制,并根据每幅图像中的可比较特征来进行检索。目前大量的基于内容的图像检索研究都是全局图像信息,然而在大多数情况下,用户更关心的是图像中具有一定语义的区域,为了达到这种效果,一些图像检索系统中引入了图像自动分割和自动区域提取技术,然而到目前为止还不存在一种通用的方法,同时也没有一个判断分割质量的标准,因而分割结果必然会导致与人的主观认识上的差异,也导致无法准确地提取相关区域的视觉特征,同时也降低了检索结果的效果。为了解决上述问题,本文主要侧重研究基于局部特征的图像检索。研究的具体的方法如下：提出了一种综合颜色和图像轮廓曲线特征的检索方法。该方法首先分割图像并提取图像中感兴趣对象的轮廓,接着对提取的轮廓进行仿射变换及最小值化处理,经处理后的轮廓带有边缘的完整信息,并具有几何不变性；其次,利用聚类的颜色信息,提取主聚类的直方图,所提取的直方图不仅包含了主聚类的颜色信息也包含了该聚类的空间位置信息。最后,利用检索对象与被检索对象的颜色距离直方图及轮廓曲线距离偏差的加权平均度量检索及被检索对象的相似性。提出了一种基于感兴趣区域的检索方法,用户需要在整幅图片中选择自己要检索的物体作为检索对象。该算法将Mean shift跟踪的思想运用到基于内容的图像检索中,但经典的Mean Shift跟踪算法利用颜色直方图来跟踪目标,并没有考虑到尺度的变化和目标像素的空间位置,针对上述问题,本文首先提出了一种快速自适应调整窗宽尺度的算法,该算法能够快速、准确的查找检索图像中目标的大小来改变窗宽的尺度,其次,在传统的直方图中加入空间信息,该信息反映目标像素的空间位置,提高跟踪的鲁棒性,最后,利用颜色分布熵来度量两幅图片的相似度,该方法更能真实反映物体的空间结构。实验表明对经典的mean shift做一系列的改进,可以提高跟踪定位的准确性和检索效果的精确性。开发了一个基于内容的图像检索引擎,该软件实现了基于内容的图像检索的主要方法。该软件基于B/S架构,软件根据图像检索的不同特征及检索需求,实现了四种不同的图像检索方法,其中实现的基于主聚类匹配的图像检索方法是本科研团队独立创新的结果,其主要优点在于该算法在提取特征时不只是局限于单一的特征提取,而是综合了图像的多重特征,从而提高了图像检索的精度。更多还原

【Abstract】 Content-based image retrieval (CBIR) is a kind of new technique, which applies image low-level information and currently existing processing methods to extract image feature and match between images. Currently most widely used CBIR methods are based on global image information. However, for practical application users are more focus on certain semantic features of an image. In order to achieve rational results, these methods based on image segmentation and auto regional feature extraction are introduced into some image retrieval systems. Unfortunately, so far there is no a general method and a segmentation quality standard. It is obvious that the human perception for vision has subjectivity, because this kind of subjectivity the effectiveness of image retrieval is affected. In order to address the above problems, this paper mainly is devoted to explore such a kind of image retrieval methods based on local feature. In this article, two kinds of new methods are proposed, i.e., an image retrieval method composited of color and shape features, an image retrieval method based on adaptive detecting and extracting an object of interesting (ROI), and an image retrieval engine system based on web has also been developed.The rest of this article is organized as follows:An image retrieval method compositing features of color and object contour curve is presented. Firstly, an image is segmented into multi-clusters. Secondly, an interesting object in image is extracted. Furthermore, its contour is extracted. Thirdly, the contour is transformed by affine, and processed by the minimum. The contour contains the whole information of interesting object and preserves geometric invariance. In addition, a histogram for primary cluster with color feature is extracted. Such an extracted histogram contains not only color information but also spatial location information. Finally, a weighted average for color distance histogram and distance deviation of contour curve is applied as similarity measure to match between two images. Experimental results show that the proposed method achieves a better retrieval precision.Another method, which is called a novel image retrieval method based on region of interest (ROI), combines multiple features with mean shift (MFMS) tracking algorithm and EM scale transformation (EMST). For typical mean shift (MS) tracking method only color histogram is considered, other features, such as spatial distribution and texture feature, are neglected. Hence, it is easy to fall failure during detection ROI. However, in our proposed method spatial distribution is integrated into MS, therefore, intuitively for detecting ROI MFMS is of a more fine effectiveness. In fact, experimental results also show that MFMS is able to detect the position of ROI more accurately and robustly than one of MS. In addition, the EMST uses EM-like algorithm to estimate the local position and covariance matrix, which describes the approximate scale of ROI. It is better to quickly and accurately describe such a scale change in the process of image retrieval.An image retrieval engine based on web has also been implemented, which is a kind of software package based on CBIR and B/S architecture. With respect to the different user, this software has implemented four different image retrieval functions. A primary clustering matching method, whose advantage is to use multi-features, is proposed by our research team.更多还原