一种基于形状上下文特征和分层聚类的快速商标识别方法.docVIP

 Fast Trademark Recognition Based on Shape Context and Hierarchical Clustering WU Qian, ZHANG Honggang, CHAI Lunshao** 5 10 15 20 25 30 35 40 (School of Information and Communication Engineering, Beijing University of Posts and Telecommunications, Beijing 100876) Abstract: Trademark recognition has been played a very important part in our daily life, which can be applied to various fields such as advertisement, logistics transportation network, E-commerce and so on. In this paper, a fast trademark recognition method based on shape context is proposed to detect and classify the trademarks in images. However, the difficulties lie in that when the sample set becomes larger, along with the calculation of the feature points matching increases, the computational efficiency greatly reduced. The method proposed in this paper attempts to cluster the feature vectors in training procedure so that the search scope of K nearest neighbors in KNN algorithm could be narrowed. The experimental results show that the modified algorithm has faster computing speed on the basis of maintaining the previous performance, and do well in trademark recognition. Key words: trademark recognition; shape context; object matching; clustering; 0 Introduction A trademark represents the identification of commodity, which is protected by the law. So trademark recognition is very important and challengeable in the field of image processing and recognition. However, it is a difficult problem to detect trademarks from a large image database. In recent years, a number of methods for trademark recognition have been proposed by domestic and foreign scholars [1], but due to differences in viewpoint, scale, affine or lighting conditions, it is still a hard problem. Generally, an image is regarded as an dimensional vector formed by contacting the brightness values of the pixel in computer vision, therefore, many kinds of object detecting methods are proposed on this representation, such as training a disc