SABADO - SmArt BrAnd DetectiOn

Abstract

The need for an automatic system that detects brands in digital content grows simultaneously with its pervasiveness and all the marketing business surrounding it. Most of the current used solutions use implicit placement of advertisements without regard for content or context, which has been shown to be an e ective way of raising brand awareness. Moreover, knowing where a brand is used (and how), could serve as support for business decisions. This work aims to explore new ways of assessing the presence of branded products or services in images using machine learning and computer vision techniques. A brand is much more than a graphic logo, it can de ned by several other subtle attributes, for example, how it is advertised and by its reputation. In this work, we tackle the problem of brand detection through the graphic logo, since it is the most tangible element of a brand. Given all the possible ways a graphic logo can appear in an image, building a highly accurate system is a challenging task. We propose a brand detection system that takes advantage of modern deep learning computer vision systems. We built several models, tuned and tested them using the FlickrLogos-32 dataset. Using the Fast Region-based Convolutional Networks (FRCNs) model our system yields a mean average precision value of 73.47%, outperforming previous methods while also having reasonably fast detection time and outputting multiple logo detections per image. From an application standpoint, many challenges still remain, but with advances recently done in computer vision and machine learning, we will come to reach a system capable to even capture subtle characteristics of a brand and help companies and brands captivate their customers and to make better decisions