Logo detection with second judge single shot multibox.

Abstract

With the increasing popularity of Social Networks, the way people interact has changed and the huge amount of data generated open doors to new strategies and marketing analysis. According to Instagram 1 and Tumblr2 an average of 95 and 35 million photos, respectively, are published every day. These pictures contain several implicit or explicit brand logos, this allows us to research how can a brand be better widespread based in regional, temporal and cultural criteria. Using advanced computer vision techniques for object detection and recognition, we can extract information from these images, making possible to understand the impact and the comprehensiveness of a specific brand. This thesis proposes a logo detection technique based on a Convolutional Neural Network (CNN), also used as a second judge. Our proposal is built on the Single Shot Multibox (SSD). In our research, we explored several approaches of the second judge and managed to reduce significantly the number of false positives in comparison with the original approach. Our research outperformed all the others researches on two different datasets: FlickrLogos-32 and Logos-32plus. On the FlickrLogos-32, we surpass the actual state-of-the-art method by 5.2% of F-score and for the Logos-32Plus by 3.0% of F-score.