بازیابی تصویر علائم تجاری با استفاده از ویژگی های مصنوعی برای توصیف شکل جهانی و ساختار داخلی

A trademark image retrieval (TIR) system is proposed in this work to deal with the vast number of trademark images in the trademark registration system. The proposed approach commences with the extraction of edges using the Canny edge detector, performs a shape normalisation procedure, and then extracts the global and local features. The global features capture the gross essence of the shapes while the local features describe the interior details of the trademarks. A two-component feature matching strategy is used to measure the similarity between the query and database images. The performance of the proposed algorithm is compared against four other algorithms.

With the rapid increase in the amount of registered trademark images around the world, trademark image retrieval (TIR) has emerged to ensure that new trademarks do not repeat any of the vast number of trademark images stored in the trademark registration system. As the traditional classification of trademark images is based on their shape features and types of object depicted by employing manually assigned codes, faults or slips may appear because of different subjective perception of the trademark images. Evidence has been provided that the traditional classification is not feasible in dealing with a large fraction of trademark images with little or no representational meanings [1]. Trademarks can be categorised into a few different types. A trademark can be a word-only mark, a device-only mark or a device-and-word mark. For a word-only mark, the design of the trademark consists purely of text words or phrases. However, for a device-only mark, the trademark only contains symbols, icons or images. If a trademark comprises both words and any iconic symbols or images, it can be regarded as a device-and-word mark [2]. Since different algorithms have to be used in describing different kinds of trademark images, a trademark image retrieval system can only be designed to accommodate one of the types. Although several trademark image retrieval systems have been designed to handle all kinds of trademark images, the performance of these systems are rather unfavourable when compared to those systems that are specifically designed to handle only one kind of trademark. Another challenge in trademark image retrieval is the difficulty in modeling human perception about similarity between trademarks. As human perception of an image involves collaboration between different sensoria, it is in fact difficult to integrate such human perception mechanisms into a trademark image retrieval system. The contributions of this paper are summarized as follows: (1) novel algorithms are proposed to describe the shape of device-only marks and device-and-word marks; (2) a two-component feature matching strategy is applied to compare global and local features; (3) this study not only evaluates the proposed method, but also investigates the retrieval performance of another four algorithms. The rest of this paper is organized as follows. Section 2 reviews the related studies regarding the existing trademark image retrieval systems and techniques. Section 3 provides an overview of the proposed system architecture. Section 4 presents the algorithms proposed for extracting global and local features of trademarks. Section 5 describes a two-component matching strategy for measuring similarity between trademarks. Section 6 evaluates the performance and analyses the results. Finally, conclusions are drawn in Section 7.

In this work, we have proposed a novel content-based trademark retrieval system with a feasible set of feature descriptors, which is capable of depicting global shapes and interior/local features of the trademarks. We have also proposed an effective two-component feature matching strategy for measure the similarity between feature sets. By utilising the curvature feature and the distance to centroid, the proposed algorithm is robust against rotation, translation, scaling and stretching. As for the image retrieval stage, a two-component matching strategy was used in feature matching. With this strategy, the images can be compared with the query image with their local and global features taken into account separately, and therefore enabling the system to be insensitive to noise or small regional changes. The performance of the proposed algorithm was evaluated in terms of the precision and recall rates. The precision–recall graphs show that the proposed algorithm outperforms other conventional algorithms, including moment invariants, the Fourier descriptors, the Zernike moments only and the CSS. Nevertheless, the proposed scheme is not sufficiently capable of relating the trademarks with similar semantic meanings but significantly different low-level features. We are currently investigating ways of incorporating relevance feedback into the proposed system in order to tackle this challenge.