| Summary: | 碩士 === 國立中正大學 === 資訊工程研究所 === 99 === Logos or trademarks are significant representation for organizations and companies. Every year, companies invest an extremely large amount of money in being famous. In daily life, logos and trademarks can be seen everywhere, such as sports games, TV series, and social network websites. If we can extract logo objects from media, we can develop many interesting applications about logos. Therefore, how to detect and localize logo in real-world images is a study of high potential commercial values.
In this research, we propose a framework to automatically detect and localize logo objects in images. Based on using local features, we utilize a pair-specific candidate features to detect approximate locations of logo objects and decrease a large number of searching time on a large-scale database. We further utilize mean-shift algorithm to localize the logo objects. In our work, a more important thing is that we consider the spatial relationships between local features to describe logo objects. We use a more descriptive feature representation, i.e., visual pattern, to describe local features configurations of logo objects. The visual pattern is more discriminative than a single local feature by modeling spatial relationships between local features, and also preserves flexibility to tackle with objects scaling, rotation, and deformation. Experimental results show that our framework has better performance than other methods, and indeed effectively detects and localizes logo objects in real-world images. Furthermore, experiments also show that considering spatial relationships between local features is superior to utilizing single feature only.
|
|---|
