容差邻域模型及其在目标遮挡的场景图像中的应用

发布时间：2018-03-05 19:40

本文选题：大数据　切入点：粗糙集　出处：《太原理工大学》2017年硕士论文　论文类型：学位论文

【摘要】：大数据处理技术的迅速发展,极大的改变了人们的生活习惯、工作方式和思维模式。专家和学者们也认识到海量数据分析和处理的广阔前景,并希望能够从中得到有用的信息。大数据往往具有不确定、维数大和不完备等特点,而粗糙集作为处理不确定、不一致问题的有效工具,已经广泛的应用在数据挖掘中。针对粗糙集常用于处理完备信息系统的问题,人们致力于寻找更好的扩展方法并将其应用于不完备信息系统。本文使用“邻域”和“容差完备度”的概念对经典粗糙集进行了扩展,得到了新的邻域模型,并将该模型应用于目标遮挡的图像分类当中。通过在Pawlak经典粗糙集的基础上引入了粗糙集的邻域模型,用于解决离散型属性和连续型属性的混合数据类型不能同时处理的问题,并介绍了邻域粗糙集(Neighborhood Rough Set,NR)的基本概念。而不具有容差能力的方法在处理不完备的信息时难以达到理想的效果,因此,本文给出了扩展容差关系(Extended Tolerance Relation,ETR)模型。该模型将限制条件设为邻域和容差完备度,并以扩展容差邻域为基础选择决策正域,经计算得到系统的属性重要度,最后给出基于扩展容差关系的属性约简算法,并通过删除冗余来降低噪声数据对分类结果产生的影响。变换不同的邻域阈值,将单个样本应用于不同的分类器,并分析实验结果。使用UCI库上的几组混合类型数据集进行实验,并与其他几种扩展粗糙集算法进行对比。通过分析各算法在不同分类器下的精度变化趋势可知,ETR算法能在不降低分类精度的同时保持较少的约简。因此验证了扩展容差邻域模型的有效性及算法的可行性。本文采用颜色和纹理相融合的方法将扩展容差关系应用在目标遮挡的场景图像分类中。首先,构建遮挡图像对象集知识表示系统,面向对象集系统,使用扩展容差邻域模型建立图像边缘及遮挡边界的相容粒度空间;其次,计算相容粒度空间下的颜色特征直方图,得到相容粒的直方图统计特征;最后,在多种对比算法下,使用不同分类器对本文算法进行验证。实验结果表明该方法在解决复杂场景图像中遮挡问题的有效性,实现了场景图像的分类与检索。
[Abstract]:The rapid development of big data's processing technology has greatly changed people's living habits, working methods and thinking patterns. Experts and scholars have also recognized the broad prospect of massive data analysis and processing. Big data often has the characteristics of uncertainty, dimension and incompleteness, and rough set is an effective tool to deal with the problem of uncertainty and inconsistency. Has been widely used in data mining. Rough sets are often used to deal with the problem of complete information systems, In this paper, the concepts of "neighborhood" and "tolerance completeness" are used to extend the classical rough set, and a new neighborhood model is obtained. The model is applied to the image classification of target occlusion. Based on the classical Pawlak rough set, the neighborhood model of rough set is introduced to solve the problem that the mixed data types of discrete and continuous attributes can not be processed at the same time. The basic concept of neighborhood Rough set (NRs) is introduced. The method without tolerance ability is difficult to achieve ideal results when dealing with incomplete information. In this paper, the extended tolerance relation extended Tolerance relation model is presented. The model sets the constraint conditions as neighborhood and tolerance completeness, and selects the decision positive domain based on extended tolerance neighborhood. The attribute importance of the system is calculated. Finally, an attribute reduction algorithm based on extended tolerance relationship is presented, and the influence of noise data on classification results is reduced by deleting redundancy. Different neighborhood thresholds are transformed, and a single sample is applied to different classifiers. The experimental results are analyzed. Several sets of mixed data sets on UCI library are used to carry out experiments. And compared with other extended rough set algorithms, by analyzing the trend of the accuracy of each algorithm under different classifiers, we can see that the ETR algorithm can keep less reduction without reducing the classification accuracy. Therefore, it is verified that the extension algorithm can reduce the classification accuracy at the same time. The validity of the neighborhood model and the feasibility of the algorithm. In this paper, the extended tolerance relationship is applied to the target occlusion scene image classification using the method of color and texture fusion. The knowledge representation system of occlusion image object set and object oriented set system are constructed, and the compatible granularity space of image edge and occlusion boundary is established by using extended tolerance neighborhood model. Secondly, the color feature histogram under consistent granularity space is calculated. The histogram statistical features of compatible particles are obtained. Finally, different classifiers are used to verify the algorithm under various contrast algorithms. The experimental results show that the proposed method is effective in solving the occlusion problem in complex scene images. The classification and retrieval of scene images are realized.
【学位授予单位】：太原理工大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【参考文献】