基于快速密度峰值聚类的图像检索技术研究与应用

发布时间：2019-05-24 08:37

【摘要】：随着人类社会的进步以及计算机、互联网、存储技术等高新科技的发展,每天都有数以亿计的图像产生并被通过各种渠道进行传播,从而导致数字图像的数量一直以一种惊人的速度在增长。对于如此庞大的图像数据,如何有效地管理和检索,并从中获取潜在的信息及价值成为了人们亟待解决的难题。因此我们需要更加快速、准确的图像检索方法来查询所需要的图像及相关信息。图像聚类为图像检索提供了新的技术支持,基于聚类的图像检索能够在大量图像数据中快速、精准的发掘用户感兴趣的信息。然而传统应用于图像聚类的特征提取算法往往忽略图像颜色的空间分布信息,且适应性较差。因此本文通过等面积矩形环对图像进行划分并计算各空间区域的相关性,并根据空间区域相关性计算各区域的重要性,将空间信息与颜色信息进行融合。同时本文研究了快速搜索密度峰值(DP)聚类算法并对其进行合理改进后运用在图像检索系统中,在保证收敛速度的同时提高了聚类精度。本文主要研究内容及工作如下:(1)研究颜色特征提取及其量化方法,常用的颜色空间一般是基于硬件角度提出的,不能很好的与人眼感知相匹配,本次研究选取HSV颜色空间作为颜色空间模型。同时为了提高运算速度,本文通过人眼对颜色的感知对其进行了量化,从而便于统计和计算。(2)研究区域相关性计算方法,传统的颜色特征提取法仅对颜色值进行统计和整理,并不考虑空间分布情况,为了使颜色特征更具代表性,本文提出一种基于图像内容的区域相关性计算方法将空间信息与颜色特征进行了融合,并根据区域相关性自动调整各区域的重要性权值,提高了特征提取算法的鲁棒性及普适性。(3)研究DP聚类算法的优化方案,原本的DP聚类算法截断距离固定不变,该参数的选取在某种意义上决定着聚类算法的成效。因此到一个合适的截断距离对DP算法聚类效果有较明显的影响。本次研究提出了一种截断距离动态调整方案,使之在保证较快的收敛速度的同时具有较高的聚类精度。最终通过实验验证,本文提出的方法是可行的、有效的。
[Abstract]:With the progress of human society and the development of high and new technology such as computer, Internet, storage technology and so on, hundreds of millions of images are produced and disseminated through various channels every day. As a result, the number of digital images has been growing at an amazing rate. For such a large image data, how to effectively manage and retrieve, and obtain potential information and value has become an urgent problem to be solved. Therefore, we need faster and more accurate image retrieval methods to query the required images and related information. Image clustering provides new technical support for image retrieval. Image retrieval based on clustering can quickly and accurately discover the information of interest to users in a large number of image data. However, the traditional feature extraction algorithms used in image clustering often ignore the spatial distribution information of image color, and the adaptability is poor. Therefore, this paper divides the image into equal area rectangular rings and calculates the correlation of each spatial region, and calculates the importance of each region according to the spatial region correlation, and merges the spatial information and color information. At the same time, this paper studies the fast search density peak (DP) clustering algorithm and improves it reasonably in the image retrieval system, which not only ensures the convergence speed, but also improves the clustering accuracy. The main research contents and work of this paper are as follows: (1) the color feature extraction and its quantification methods are studied. The commonly used color space is generally based on the hardware point of view, which can not match the human eye perception very well. In this study, HSV color space is selected as color space model. At the same time, in order to improve the operation speed, this paper quantifies the color perception by the human eye, so as to facilitate statistics and calculation. (2) the regional correlation calculation method is studied. The traditional color feature extraction method only statistics and collates the color value, and does not consider the spatial distribution, in order to make the color feature more representative. In this paper, a regional correlation calculation method based on image content is proposed to merge spatial information with color features, and the importance weights of each region are automatically adjusted according to the regional correlation. The robustness and universality of the feature extraction algorithm are improved. (3) the optimization scheme of DP clustering algorithm is studied. the truncation distance of the original DP clustering algorithm is fixed, and the selection of this parameter determines the effectiveness of the clustering algorithm in a sense. Therefore, reaching a suitable truncation distance has obvious influence on the clustering effect of DP algorithm. In this study, a dynamic adjustment scheme of truncation distance is proposed, which not only ensures the fast convergence speed, but also has high clustering accuracy. Finally, the experimental results show that the method proposed in this paper is feasible and effective.
【学位授予单位】：重庆理工大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【参考文献】