基于空间关系相似性的图像检索

发布时间：2018-04-23 08:36

本文选题：相关反馈 + 定性空间推理　；参考：《吉林大学》2012年硕士论文

【摘要】：随着互联网技术的高速发展、各种形式的媒体不断增加、互联网上以图像形式展现的内容越来越多，图像检索逐步成为了主要的检索方式之一。互联网上传统的搜索引擎，例如Google、Yahoo！、Bing等提供的图像检索，都是基于图像的文件名、所在网页内嵌的关键字等机制施行的查询。由于图像的理解与文本表达之间的鸿沟，基于关键字的图像检索的准确率不高，因此，研究者们提出了基于内容的图像检索(Content-BasedImage Retrieval，CBIR)。基于内容的图像检索是指查询条件本身就是一幅图像，或是对图像内容的描述，通过提取底层或者高层的特征建立索引，计算这些特征和查询特征之间的距离，得到两幅图像间的相似程度。基于空间关系相似性的图像检索属于CBIR的一个分支。该方法在获取对象的类别之后，找出在两个图像内均出现的对象，将这些对象组成一个对象集合，判断该对象集合内哪些对象在两个图像中构成的空间关系是一致的，将空间关系一致的对象标志为“相似对象”。进行图像检索时，用户有时无法清楚描述目标图像的特征，提供的查询图像只具有目标图像的部分特征，这种情况下，系统需要在第一次推荐之后进行多次相关反馈。通过用户在推荐结果中选择若干与目标图像相关的结果（即正例），利用这些正例中公共的对象、以及公共对象间一致的空间关系等有效信息使目标图像的特征描述不断趋于精确，最终检索系统可根据较精确的查询特征提供目标图像。本文的研究工作分为两部分：首先，本文通过对公共模式方法(Common Pattern Method，CPM)的改进，提出了基于矩形代数的相似性图像检索(Similarity Retrieval by Rectangle Algebra，SRRA)算法。SRRA算法将对象抽象成最小边界矩形，利用矩形代数判断对象间的空间关系。和基于CPM的算法所采用的基于点的空间关系模型相比，最小边界矩形能更精确地表示出对象的空间大小、范围等信息，而且利用矩形代数表示对象间的空间关系比CPM中采用的type-i方法更严格，使得最终的检索结果更加精准；此外，由于检索过程中剪枝掉了大量冗余的对象，SRRA算法能显著减少检索时间。其次，本文采用查询特征重定义的方法，提出了基于SRRA的相关反馈处理算法。该算法首先将每幅图像转化为一个有序符号串，针对所有正例的符号串求得一个最长序列模式，该最长序列模式表示所有正例包含的公共对象、以及公共对象间一致的空间关系；其次求得最长序列模式与查询图像的符号串的最短公共超串，作为新查询图像的符号串；最后利用矩形代数基本关系复合、路径相容原理相结合的方法，计算新的符号串中对象间的矩形代数关系。经过以上三个步骤，对查询特性进行了重定义，进入到下一轮反馈；当反馈到一定次数，或者由用户确认终止反馈时，检索结束。本文采用C++与matlab混合编程的方式实现了基于空间关系相似性的图像检索系统。同时，针对以上两方面的研究，本文比较了SRRA算法与CPM的检索时间和检索效果，给出了基于SRRA的相关反馈处理算法的实验结果。本文的实验采用自动生成的图像数据和来自http://wang.ist.psu.edu/docs/related/的数据库，实验结果表明： 1.与CPM相比，SRRA算法不但提高了检索结果的精确性，而且显著缩短了检索时间，其运行时间比CPM减少了近50%； 2.基于SRRA的相关反馈处理算法在原始查询图像的特征基础上，通过提取正例中的公共对象、公共对象间一致的空间关系等有效信息，不断精化查询图像，使得系统能更清晰地刻画目标图像的空间特征，从推荐结果中不断删除无关图像，并最终将符合目标特征的结果推荐给用户。
[Abstract]:With the rapid development of Internet technology, various forms of media are increasing, and more and more images are displayed on the Internet, and image retrieval has gradually become one of the main retrieval methods. The traditional search engines on the Internet, such as Google, Yahoo!, and Bing, are all images based on the name of the image file. Because of the gap between image understanding and text expression, the accuracy of keyword based image retrieval is not high because of the gap between image understanding and text expression. Therefore, the researchers proposed Content-BasedImage Retrieval (CBIR) based on content. Content based image retrieval means the query condition itself It is an image, or a description of the content of the image. The index is established by extracting the features of the bottom or the high level, and the distance between these features and the query features is calculated, and the similarity between the two images is obtained.
Image retrieval based on spatial relation similarity belongs to a branch of CBIR. After obtaining the category of the object, this method finds out the objects that appear in the two images, and makes up a set of objects, and judges which objects in the set are in the same spatial relationship in the two images. In the case of image retrieval, the user sometimes can not clearly describe the feature of the target image. The query image provided only has some features of the target image. In this case, the system needs to carry out multiple correlation feedback after the first recommendation. The result of the target image (that is, the positive example), using the common objects in these positive examples, and the consistent spatial relationship between the public objects and other effective information, the feature description of the target image tends to be accurate, and the final retrieval system can provide the target image according to the more accurate query features. The research work of this paper is divided into two parts:
First, through the improvement of Common Pattern Method (CPM), this paper proposes a rectangular algebra based similarity image retrieval (Similarity Retrieval by Rectangle Algebra, SRRA) algorithm.SRRA algorithm to abstract the object into a minimum boundary rectangle, and uses rectangular algebra to determine the spatial relationship between objects. Compared with the point based spatial relation model, the minimum boundary rectangle can more accurately represent the space size, range and other information of the object, and the spatial relation between the objects is more strict than the type-I method used in CPM by the rectangular algebra, which makes the final retrieval result more accurate; in addition, the scissors in the retrieval process are cut. A large number of redundant objects are dropped, and SRRA algorithm can significantly reduce retrieval time.
Secondly, this paper proposes a SRRA based correlation feedback processing algorithm based on query feature redefinition. This algorithm first transforms each image into an ordered string, and the longest sequence pattern is obtained for all the positive examples. The longest sequence pattern represents all the common objects, as well as the public objects, which are included in the positive examples. The shortest common superstring of the longest sequence pattern and the symbol string of the query image is obtained as the symbol string of the new query image. Finally, the rectangular algebraic relationship between the new symbol strings is calculated by using the method of combining the basic relation of the rectangular algebra and the path compatibility principle. The above three steps are carried out. In the end, the query feature is redefined and entered into the next round of feedback; when the feedback reaches a certain number, or the user confirms the termination of feedback, the search ends.
In this paper, the image retrieval system based on spatial relationship similarity is realized by using C++ and MATLAB hybrid programming. At the same time, for the above two aspects, this paper compares the retrieval time and retrieval effect of SRRA algorithm and CPM, and gives the experimental results of the correlation feedback processing algorithm based on SRRA. The experiment of this paper is automatically generated. Image data and database from http://wang.ist.psu.edu/docs/related/, the experimental results show that:
1. compared with CPM, SRRA algorithm not only improves the accuracy of retrieval results, but also significantly reduces retrieval time, and its running time is reduced by nearly 50% compared with CPM.
2. based on the characteristics of the original query image, the SRRA based correlation feedback processing algorithm constantly refined the query images by extracting the common objects and the consistent spatial relations among the public objects, making the system more clearly depicting the spatial features of the target image and continuously deleting the unrelated images from the recommended results. And finally recommend the results that conform to the target characteristics to the users.

【学位授予单位】：吉林大学
【学位级别】：硕士
【学位授予年份】：2012
【分类号】：TP391.41

【参考文献】