The Technique Of Trademark Image Retrieval Based On Multi-feature Combination

At presents, trademark which is a product of the development of commodity economy, is one of the most important commodity logos. It's used to distinguish products or service provided by different producers and to protect their benefits. So far, trademark is playing a more and more important role in our economical life. As we all know, trademark retrieval can prevent Unscrupulous businessmen or individuals, who want to obtain illegal benefits, to register by similar or counterfeit trademarks in legal ways, thereby protect the interests of legitimate businessmen. Meanwhile, it's hard to retrieve those trademarks which involve graphic trademarks, combined trademarks and the trademarks absent from complete information by text. However, content-based trademark image retrieval is widely used to resolve the problems and is a popular application currently.In order to improve its singularity and accuracy of the current retrieval system, this paper has built an experimental retrieval system of multi-feature combination based on contents, with classification and feedback functional modules. The research involves how to classify single target sub-images of binary trademark images, a number of target sub-images of binary trademark images and colorful trademark image, and how to segment different classifications of trademark images with Canny and Region-growing operators. For single target sub-images of binary trademark images, shape-region and shape-contour descriptors are extracted; for a number of target sub-images of binary trademark images, feature vectors based on spatial angle relationships and shape are used to describe the trademark images; for colorful trademark images, color moments, color histogram and shape-region descriptors can express the colorful trademark images. Gaussian normalization is used to normalize and integrate the different feature vectors and absolute Euclidean distance similarity algorithm is applied initially to retrieve the trademark images. What's more, this experimental system adopts a feedback module and users can estimate the results, adjust the weights needed and retrieve again. It's proved that the retrieval results got by the way of multi-feature combination are much better than ones retrieved by single features, and weights adjusted by user feedback can retrieve better results.There are mainly three points in this paper. Firstly, a classification module is used to sort different trademark images. Secondly, a feedback module is adopted to improve the way of weights of color feature vectors and good results are obtained. Thirdly, a multi-feature combination based on contents of experimental system is designed and developed.