The Trademark Image Retrieval Based On Multiple Feature Fusion

Recently, with the rapid increasing number of registered trademarks, it is urgently necessary to establish an accurate and efficient trademark image retrieval system, which is used to avoid similar trademarks registered repeatedly and spitefully. Now, content-based image retrieval is most widely used in the field of trademark retrieval system. The technology extracts features and matches features automatically by computer and ultimately returns images which are relevant to the target image. Not only does this method save a lot of human work and time, but also the retrieval result can satisfy visual perception of human more.Now, none CBIR trademark retrieval system can full fill command of different users, study and improvement of technologies which are used for trademark image retrieval are still needed.For binary trademark images, the retrieval system is usually developed by the shape characteristic because the color information and texture features are not abundant. Moments feature is an important shape feature which can describe image effectively. Generally speaking, moment feature has good stability to deformation、rotation and illumination and so on. Being used in the field of trademark image retrieval, moment feature can made accuracy retrieval result. So, it’s valuable to study and make some improvement on the basis of existing research. In this article, I will improve the method of similarity comparison of Zernike Moments and Psedudo-Zernike Moments to improve the performance of the retrieval system.Due to the advantage of multi-feaure fusion, the improvement refered above will be fused with other shape descriptors. On the basis of multi-feaure fusion, a trademark image retrieval system will be designed and developed. To enlarge the scope of application, the system will apply many feature fusion schemes for different kinds of trademark images. Besides, to solve the problem that the retrieval results of few images are not satisfied, the system can make further retrieval with SIFT descriptors when users’feedback are not satisfied. The experiment result shows that the effect of retrieval of the system conforms to the human visual sense more.