CASIA OpenIR  > 毕业生  > 硕士学位论文
Alternative TitleStudy of Accurate Image Retrieval Combined with SIFT and Color-Invariant
Thesis Advisor肖柏华
Degree Grantor中国科学院研究生院
Place of Conferral中国科学院自动化研究所
Degree Discipline模式识别与智能系统
Keyword颜色不变性 Sift 图像精确检索 Fast Adaptive Meanshift Clustering Color-invariant Sift Image Accurate Retrieval Fast Adaptive Meanshift Cluster
Abstract随着信息时代特别是互联网和多媒体技术的到来,人们每天都会被各种各样的图像信息所充斥着。因此,从纷扰复杂的图像中提取自己所需要的信息就成为人们所非常关注的话题,同时这也是一个非常富有挑战性的研究课题。 现存的图像检索系统采用的方法主要可以分为两类:第一类就是基于SIFT特征的聚类方法,采用向量空间模型和倒排索引的快速检索方法; 另一类就是基于颜色、纹理、形状等全局特征,采用直方图相交等的快速匹配算法。第一类方法具有很好的鲁棒性,是采用图像的局部特征进行检索的方法,它有效地降低了图像旋转、平移、扭曲等对检索效果造成的不良影响。另一类方法是根据图像的全局特征,这种方法简洁、高效,特征的维数也比较低。但是,全局特征会将图像的前景和背景融为一体,很难满足用户精确查找的要求。 为了克服上述两种方法的缺点,本文采用两种方法将图像的颜色特征和SIFT特征相融合来提高图像的检索精度。第一种方法是首先提取图像的SIFT关键点并将其定位到二维图像中,然后将关键点处的RGB颜色分量表示为具有颜色不变性空间的颜色分量,并建立颜色不变性空间的颜色直方图,最后将得到的频率值作为向量空间检索模型中的tf-idf值。第二种方法是直接获取整幅图像的颜色不变性空间的直方图并将其作为tf-idf值。实验结果表明第一种融合方法具有更好的检索精度。 为了克服高维聚类的性能瓶颈,本文分析了SIFT特征的提取过程和实验结果,发现了128维的SIFT特征具有分布不均匀的特点,而MeanShift聚类算法的性能瓶颈最后可以归结为寻找k-近邻的性能问题,因此,本文在总结分析现有快速k-近邻算法的基础上,提出了一种适合高维SIFT数据的 Adaptive LSH算法,并将其应用于图像检索Bag of Features模型的码本聚类中,有效地提高了模型生成速度。 在上述研究工作的基础上,我们开发了一个基于内容的图像检索实验系统,并给出了一种基于手工绘制目标草图来进行检索的设想。
Other AbstractWith the information age, especially the advent of Internet and multimedia technology, we will confront with a wide variety of images every day. Therefore, how to extract useful information from the turbulent images becomes a very concerned topic. Current image retrieval system can be divided into two categories: the first one is based on SIFT feature clustering method, and then using vector space model and fast inverted index retrieval method; the second category is based on low-dimensional features, for example color, texture, shape features, and then using histogram intersection to realize the fast matching algorithm. The first method is robust, and uses the local features to retrieval the query image, which effectively reduces the defects caused by the image rotation, translation and twist transformation. The second type is based on the image global features, this method is simple, efficient and feature dimension is relatively low. But global features will blend the foreground and background together; it is difficult to meet accurate requirements of uses. To overcome the former two defects we propose two different fusion methods between image’s SIFT and color features to increase the retrieval accurate, in this paper. The first one is that, firstly we extract image’s SIFT key points and relocate it to the two dimensional image, and then transform its RGB color space to the color invariant space, secondly we create the histogram of the invariant space, lastly we use the frequency value as the tf-idf in the vector space model. The second one is that, we create the color invariant histogram directly and use the frequency value as the tf-idf. Our experiment results demonstrate that the first fusion method is the better one. To decrease the complexity of the higher dimensional cluster, in this paper, we find that the 128 dimensional SIFT feature is not well-distributed from the analysis of the feature extract process and the experiment results. Therefore, this paper analyzed the existing fast k-nearest neighbor algorithm based on a suitable for high dimensional data, Adaptive LSH SIFT algorithm, and applied to image retrieval Bag of Features Model code of the clustering; that effectively improved the speed of model clustering. On the basis of the above study, we developed a content-based image retrieval experimental system, and present a target based on hand-drawn sketch ideas to be retrieved.
Other Identifier200728014628089
Document Type学位论文
Recommended Citation
GB/T 7714
崔玉征. 基于SIFT和颜色不变性的精确图像检索算法的研究[D]. 中国科学院自动化研究所. 中国科学院研究生院,2010.
Files in This Item:
File Name/Size DocType Version Access License
CASIA_20072801462808(2551KB) 暂不开放CC BY-NC-SAApplication Full Text
Related Services
Recommend this item
Usage statistics
Export to Endnote
Google Scholar
Similar articles in Google Scholar
[崔玉征]'s Articles
Baidu academic
Similar articles in Baidu academic
[崔玉征]'s Articles
Bing Scholar
Similar articles in Bing Scholar
[崔玉征]'s Articles
Terms of Use
No data!
Social Bookmark/Share
All comments (0)
No comment.

Items in the repository are protected by copyright, with all rights reserved, unless otherwise indicated.