Grouping Local Features For Logo Recognition

As a symbol of corporations, products, sport leagues, countries etc., logos play a very important role in people’s daily lives. So automatic logo recognition technology can lead to many meaningful applications, such as the products recommendation system used in E-business and the detection system of advertisement exposure rate. O n the other hand, logo recognition as a part of image understanding task has raised interests of a large number of researchers all over the world.Object recognition is a long-standing problem in computer vision community. Although there are some constraints in logo recognition which is a sub-problem of object recognition, it’s more difficult of logo recognition than object recognition in general due to various reasons. Typically logos occupy only a relatively small part of the images. Moreover there may be several forms of the same logo. Therefore, extracting effective description of logo is a critical problem in logo recognition.In this paper we have made a thorough survey on existing logo description methods and discussed the merits and demerits of several local and global image features and feature bundling methods. As the existing method can not work well in the situation of large logo pose changes, we propose a new logo recognition method which can efficiently bundle local feature together. As the mismatch rate of local feature descriptors is very high when the logo images match to a large database, we propose a new local feature grouping method. We use the generalized Hough Transform to build a spatial mapping of local features between query image and reference image. Our method adds spatial information of local features to the traditional bag-of-word model by bundling local feature together. Compared with other feature bundling methods out method is more efficient and flexible. In order to overcome the problem of high mismatch rate caused by different forms of the same logo, we use the similarity scores of testing image and reference image as the representation of testing image. Finally, we use a linear SVM classifier as the decision-making model.Experiments on public available dataset show that our method work well in real world images and achieves state-of-the art performance.