Logo recognition for mobile advertisement

Abstract

This Final Year Project report covers the theoretical background, exploration of various aspects, and performance analysis of a logo recognition system which targets to enable mobile advertisement application. Two techniques were proposed to address the prominent problem of background noise.

The project was based on the Bag-of-Words framework where each image is modeled as a collection of features. To simulate the user scenarios that users are to capture the logo images with mobile devices centered at the logo and placed right on top, the database construction follows certain criteria.

Two databases were constructed in the project, one preliminary database for initial exploration and one final database of 30 logo subjects for system performance analysis. The selection of feature extraction technique for BoW framework, keypoint or dense sampling, was also discussed. A logo recognition scheme called Keypoint Match Ranking was developed to facilitate understanding of the robustness of BoW framework, especially SIFT descriptor and classifier.