Exemplar-based logo and trademark recognition

Automatic logo recognition facilitates the process of searching and analyzing digital image-based resources. This is due to the simplicity and informativity of logos which can be used as a visual or semantic clue for applications such as content-based image retrieval. For automatic shape recognition, model-based approaches have gained a lot of attention due to their theoretical background and promising results. However, one of the major challenges of these approaches for logo recognition is collecting samples of logos with possible perspective tilts and non-rigid deformations that most likely happen to natural images. Therefore, a robust model cannot be trained due to the lack of sufficient samples. In this paper, we aim to overcome, or at least alleviate, this challenge by introducing an exemplar-based approach. The proposed method synthesizes new images of each logo of interest, so-called exemplars, by simulating two camera axes. This is to produce almost all possible views of a logo that may appear in a given image and guarantee that most of the distortions in the shape of a logo caused by a variation of the camera optical axis direction are covered. Subsequently, an exemplar-based model is derived based on all the exemplars using linear SVMs. Finally, for the detection and recognition, we associate observed logos with the exemplars using our exemplar-based model. Experiments are carried out on two publicly available logo datasets: FlickrLogos-27 and FlickrLogos-32. The results show the superior detection and recognition capabilities of the proposed method over the competitors.