A Comparison of methods for trademark retrieval in a large trademark dataset

Tuerxun, Wusiman
By the end of the first decades of the 21 th century, the applications for trademarks worldwide have approached an astounding number of 5 million. In Turkey alone, this number has reached the 1 million mark, and is expected by all means to keep increasing. The accelerating competition of trademark influence and uniqueness has intensified trademark piracies and infringements, and therefore resulted in a substantial burden for the patent offices, not to mention direct economic losses. To overcome this ever-increasing problem of trademark registration without compromising from service quality, and to minimize unnecessary disputes of legal ownership, organizations like patent offices gradually turn to automatized registration, employing Trademark Retrieval Systems (TRS) equipped with image processing and computer vision tools. The last two decades have seen the successful implementation of well-known content-based image processing (CBIR) techniques in several TRS systems. However, these results are falling behind as the rapid increase in trademark registration escalates the trademark retrieval problem to the next level. Developing next-generation TRS systems with well-examined and analyzed new image retrieval and object detection techniques is necessary. Yet, the lack of public large-scale trademark datasets has obstructed the progress of this research field. In this thesis, to fill this gap, we offer a large scale and challenging dataset with 1 million trademarks as a benchmark. Then, as an initial attempt to trademark retrieval research on large scale datasets, we implement and analyze a variety of global image descriptors (\eg, color, shape, texture), as well as local image descriptors (\eg, SIFT, SURF, HOG, ORG), on this dataset (called the METU trademark dataset).
Citation Formats
W. Tuerxun, “A Comparison of methods for trademark retrieval in a large trademark dataset,” M.S. - Master of Science, 2015.