基于二进制分辨矩阵的视觉单词约简方法研究

发布时间：2018-04-21 06:00

本文选题：场景分类 + 视觉词包　；参考：《太原科技大学》2016年硕士论文

【摘要】：随着大数据时代的来临,互联网上图像数据正在爆炸式增长,面对越来越多的图像数据,传统的人工方式标注图像已无法满足实际需求,寻找一种快速自动标注图像方法成为需要研究的主要内容之一。目前,图像场景分类是图像语义自动标注的一个研究热点,视觉词包模型是图像场景内容表达的一种重要方式,但由于视觉词包模型形成过程中会产生冗余视觉单词、“多义词”和“同义词”,这些视觉单词的存在严重影响了图像场景的分类性能。二进制分辨矩阵方法是粗糙集属性约简中的一种有效方法,本文将二进制分辨矩阵与视觉词包模型相结合,对视觉单词的约简和场景分类方法进行了研究,主要的研究内容如下:(1)给出了一种基于二进制分辨矩阵的冗余视觉单词约简方法。该方法首先通过调整归一化阈值α的取值,对所有训练图像产生不同的0-1信息决策表和构造不同的二进制分辨矩阵;然后以二进制分辨矩阵行列方向上1的个数作为启发信息识别核视觉单词和重要视觉单词;并以这些视觉单词作为描述图像场景分类的决策规则,从而减少了冗余视觉单词对图像场景分类的影响,进而提高了图像场景分类精度。最后在OT库8类图像数据集上进行实验,验证了该方法是有效的。(2)给出了一种基于二进制分辨矩阵的多义视觉单词约简方法。由于在(1)方法中归一化阈值α的选取对决策规则的生成影响较大,而且随着视觉单词容量增大,删除视觉单词过多导致决策规则区分力度下降,因此针对任意两类不同训练图像形成0-1信息决策表并构建二进制分别矩阵;然后根据二进制分辨矩阵约简算法,将其中一类图像分别与其它不同类图像的约简视觉单词求并集运算,并以这个并集作为决定这一类图像的决策规则,从而减少了任意两类图像视觉词包中存在的“多义词”问题,进而形成区分能力更强的决策规则。最后在OT库8类和Fei-Fei Dataset 13类图像数据集进行实验,验证了该方法的有效性。(3)开发了一个基于二进制分辨矩阵的图像场景分类原型系统。基于研究内容(1),以Matlab和Java作为开发工具,设计并实现了一个基于二进制分辨矩阵的图像场景分类原型系统。
[Abstract]:With the advent of big data era, the image data on the Internet is increasing explosively. In the face of more and more image data, the traditional manual method can no longer meet the actual needs. Finding a fast and automatic image tagging method has become one of the main research topics. At present, image scene classification is a hot topic in image semantic automatic tagging. Visual word packet model is an important way to express the content of image scene, but in the process of forming visual word packet model, redundant visual words will be produced. The existence of "polysemy" and "synonym" seriously affects the performance of image scene classification. Binary resolution matrix method is an effective method in attribute reduction of rough set. This paper combines binary discernibility matrix with visual word packet model to study the reduction of visual words and the method of scene classification. The main research contents are as follows: (1) A redundant visual word reduction method based on binary resolution matrix is presented. Firstly, by adjusting the value of normalized threshold 伪, the method produces different 0-1 information decision tables for all training images and constructs different binary resolution matrices. Then, the number of 1 in the column direction of binary discernment matrix is used as the heuristic information to recognize the visual words and important visual words, and these visual words are used as the decision rules to describe the classification of image scene. Thus, the influence of redundant visual words on image scene classification is reduced, and the accuracy of image scene classification is improved. Finally, an experiment is carried out on 8 kinds of image data sets in OT library, which proves that this method is effective.) A polysemous visual word reduction method based on binary resolution matrix is presented. Because the selection of normalized threshold 伪 has a great influence on the generation of decision rules, and with the increase of visual word capacity, too much deletion of visual words leads to the decrease of decision rule differentiation. Therefore, the 0-1 information decision table is formed for any two kinds of different training images and binary separate matrix is constructed. Then, according to the binary resolution matrix reduction algorithm, One of the images is merged with the reduced visual words of the other kinds of images, and the union is taken as the decision rule to determine this kind of image. Thus, the problem of polysemous words in any two kinds of image visual word packets is reduced, and a more discriminative decision rule is formed. Finally, an image scene classification prototype system based on binary resolution matrix is developed by experiments on 8 classes of OT database and 13 classes of Fei-Fei Dataset data sets, which verify the effectiveness of this method. A prototype image scene classification system based on binary resolution matrix is designed and implemented based on Matlab and Java.
【学位授予单位】：太原科技大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TP391.41

【相似文献】