外观专利图像分类方法研究

发布时间：2018-05-15 18:28

本文选题：外观设计专利图像 + 图像视觉特征　；参考：《广东工业大学》2013年硕士论文

【摘要】：随着知识经济和经济全球化深入发展,知识产权日益成为国家发展的战略性资源和国际竞争力的核心要素。外观设计专利作为知识产权的一项重要内容,我国政府、企业对外观设计专利的保护越来越重视。通常专利图像数据库都是海量的,因此开发并发展基于内容的外观专利图像检索系统是十分必要的,同时具有深远、重大的社会、经济效益。外观专利图像检索系统在检索过程中往往只是简单的比较图像视觉特征之间的相似度,并没有按语义检索图像。而且图像库中的图像通常是海量的,顺序检索的计算量十分巨大,也是十分耗时的。针对以上问题,将这些图像划分为一些有意义的类别成为越来越迫切的需求,即实现自动分类。自动分类不但能满足用户根据图像语义内容检索的要求,还能提高检索速度。因此,图像根据语义分类是一个值得深入研究的领域。本文以外观专利图像的边缘轮廓距离作为基础数据,在兼顾外观专利图像语义相似和低层特征相似时,分别使用支持向量机(SVM,Support Vector Machine)、K均值聚类、NJW谱聚类对外观专利图像分类,并提出一种基于均值的谱聚类特征向量选择算法。针对上面四种分类算法,设计了一整套实验方案用来外观专利图像分类。实验表明,当图像库的数据量较小时,四种算法的分类效果较差,但随着数据量的增大,分类准确率得到明显的改善,并趋于稳定的状态。在简要介绍外观专利检索技术和图像分类方法现状的基础上,论文主要做了以下三个方面工作： (1)阐述了支持向量机的基本思想和分类器的构造,并将外观专利图像特征数据作为分类器的输入,实现自动分类。 (2)在兼顾外观专利图像语义相似和低层特征相似时,介绍使用K均值聚类算法实现外观专利图像分类的步骤。 (3)介绍了谱聚类的基本原理和实现步骤,提出基于均值的谱聚类特征向量选择算法,并将外观专利图像特征数据作为试验的数据集,验证K均值聚类算法、NJW谱聚类算法和基于均值的谱聚类特征向量选择算法在该数据集上分类的有效性。同时在相同特征数据的情况下,分析了不同分类方法对图像分类效果的影响。
[Abstract]:With the development of knowledge economy and economic globalization, intellectual property has become the strategic resource of national development and the core element of international competitiveness. As an important part of intellectual property, our government and enterprises pay more and more attention to the protection of design patents. The patent image database is usually massive, so it is necessary to develop and develop the content-based appearance patent image retrieval system, which has far-reaching, significant social and economic benefits. In the process of image retrieval, the appearance patent image retrieval system only simply compares the similarity between the visual features of the image, and does not retrieve the image according to the semantics. Moreover, the images in the image database are usually massive, and the computation of sequential retrieval is very large and time-consuming. To solve the above problems, it is more and more urgent to divide these images into some meaningful categories, that is, to realize automatic classification. Automatic classification can not only meet the requirements of image semantic content retrieval, but also improve the retrieval speed. Therefore, image classification based on semantics is an area worthy of further study. In this paper, the edge contour distance of patented appearance image is taken as the basic data. When semantic similarity and low-level feature similarity are taken into account, support vector machine (SVM) support Vector machine is used to classify the patented appearance image by NJW spectrum clustering. A spectral clustering feature vector selection algorithm based on mean value is proposed. Aiming at the above four classification algorithms, a set of experimental schemes are designed for image classification of patented appearance. The experimental results show that the classification effect of the four algorithms is poor when the amount of data in the image database is small, but with the increase of the amount of data, the classification accuracy is obviously improved and tends to be stable. On the basis of a brief introduction to the present situation of patent retrieval technology and image classification methods, the thesis mainly does the following three aspects: In this paper, the basic idea of support vector machine and the construction of classifier are expounded, and the feature data of appearance patent image are taken as the input of classifier to realize automatic classification. This paper introduces the steps of applying K-means clustering algorithm to realize the classification of patented appearance images by taking into account the semantic similarity and low-level feature similarity of appearance patent images. In this paper, the basic principle and implementation steps of spectral clustering are introduced, and a spectral clustering feature vector selection algorithm based on mean value is proposed, and the feature data of patented appearance image is taken as the experimental data set. The validity of the K-means clustering algorithm, NJW spectral clustering algorithm and the mean-based spectral clustering feature vector selection algorithm on the data set is verified. At the same time, the effect of different classification methods on image classification is analyzed under the same feature data.
【学位授予单位】：广东工业大学
【学位级别】：硕士
【学位授予年份】：2013
【分类号】：TP391.41

【参考文献】