Elliptical ASIFT Agglomeration in Class Prototype for Logo Detection

Raluca Boia, Corneliu Florea and Laura Florea

Abstract

Logo localization and recognition is difficult in natural images due to perspective deformations, varying background, possible occlusions, scaling variability; furthermore, the re-branding induces changes in the logo color palette and spatial distribution. To address this task, we locate keypoints using Affine Difference of Gaussian described by SIFT elliptical features; we construct the class prototypes by analyzing the graph of homographic matching between examples of the same class. The interconnections graph is developed for each class and the representative points from the non central examples are added to the class model. Potentially, an inverted secondary model is built for classes containing color inverted logos. Finally, each class is depicted by the reunion of the suitable keypoints and descriptors. The logo integrated detection (localization and classification) system is tested on multiple databases leading to state of the art accuracy.

Session

Poster 2

Files

Extended Abstract (PDF, 438K)
Paper (PDF, 4M)

DOI

10.5244/C.29.115
https://dx.doi.org/10.5244/C.29.115

Citation

Bibtex

@inproceedings{BMVC2015_115,
	title={Elliptical ASIFT Agglomeration in Class Prototype for Logo Detection},
	author={Raluca Boia and Corneliu Florea and Laura Florea},
	year={2015},
	month={September},
	pages={115.1-115.12},
	articleno={115},
	numpages={12},
	booktitle={Proceedings of the British Machine Vision Conference (BMVC)},
	publisher={BMVA Press},
	editor={Xianghua Xie, Mark W. Jones, and Gary K. L. Tam},
	doi={10.5244/C.29.115},
	isbn={1-901725-53-7},
	url={https://dx.doi.org/10.5244/C.29.115}
}