Research On Classification Methods Of Design Patent Images

With the further progress of knowledge economy and economic globalization, intellectual property is becoming a strategic resource for the country’s development and the core element of international competitiveness. As the patent for design is an essential element of intellectual property, so our government and enterprises pay more and more attention to protect the patent for design. The databases of design patent image are usually massive, so that it is necessary to develop and provide more power to the system that is Content-Based design patent Image Retrieval(CBIR),and also has far-reaching and significant social and economic benefits.Image retrieval system on design patents often simply compare the similarity of image visual features, and do not take into account the semantic features of images during the retrieval process. It is usually a mass of images in the image library, so that it will take huge amounts of computational in sequence retrieval and is very time-consuming. Faced with the above problems, dividing them into some meaningful categories become more and more pressing needs, namely, automatic classification. Automatic classification will not only can satisfy the user according to the requirements of image semantic retrieval, but also improve the retrieval speed.In this thesis, we use the bounding box-contour distances(BBCD features) of appearance design patent image as a basic data. Considering the semantic similitude and the low-level visual features similitude, we make classification on appearance design patent image, using the K-means clustering algorithm, NJW spectral clustering algorithm, support vector machine(SVM) algorithm respectively and propose an algorithm of eigenvector selection for spectral clustering based on mean. Against the above four kinds of classification algorithms, designed a set of experiments used the appearance design patent image classification. Experiments show that when the image data is low, four kinds of classification algorithms were less effective, but as the volume of data increases, classification accuracy can be significantly improved and stabilized state.On the basic of briefly introduced the current situation of appearance patent retrieval technology and image classification, this thesis has mainly the following three aspects: (1) Described the basic ideas of support vector machine(SVM) and classifier construction, and the appearance patent image data as the input of the classifier, to achieve automatic classification.(2) Considering the semantic similitude and the low-level visual features similitude, we introduce how to use the method of K-means clustering algorithm to implement the steps of appearance design patent image classification.(3) Introduced the basic principles of spectral clustering and implementation steps. and proposed an algorithm of eigenvector selection for spectral clustering based on mean(ESSCBM). We make the appearance patent image data as test data sets, and verify K-means clustering algorithm, NJW spectral clustering algorithms and ESSCBM algorithm in the data set on the validity of the classification. At the same time, we analyze the impact of the effects of different classification methods for image classification in the case of the same data.