Brand Recognition in Online Display Advertising : An Exploratory Study Based on Logo Detection

Abstract: While the problem of automatic brand detection has already been studied for in-the-wild image inputs, its resolution in the context of online display advertising remains unexplored. Following the existing works that have successfully applied logo detection techniques to photographs of the real world, this thesis builds upon modern object detection models to perform brand-agnostic logo detection in images of online advertisements. Inference with the resulting convolutional neural network allows the development of image-based brand recognition pipelines that scale to a large number of identifiable brands. Furthermore, to train and evaluate such a network, a logo detection dataset composed of captures of online advertisements is created. Each sample image is annotated with bounding boxes and brand labels. Finally, with the aim of introducing a different approach to the problem, we also design and analyze the performance of a brand detection algorithm based on the extraction and filtering of domain names from the HTML code underlying online advertisements.