Overview / Usage

I. Introduction
Advertisements have become ever more present in daily life. Google and Facebook have built major businesses selling data for advertisements. Even before those companies earned billions there had been advertising on the internet. Firefox first gained prominence by including a built in popup blocker to fight against ads, but that did nothing to stop in page advertisement. For nearly a decade there has been a way to fight back against the visual clutter brought by constant in your face advertisements on many website, Ad Block. The experience of the internet is nearly a night and day difference. Since Ad Block there have been a number of improvements, such as Ad Block Plus, uBlock Origin, and Ad Nauseum.
However currently there exist no way of blocking advertisements and the visual clutter they bring from real life. There have been some local efforts, such as the Sao Paulo law that bans outdoor advertisements, but there have not been any large scale efforts. I am proposing a first step in that process. For my project I would like to build a system that detects logos in images and erases them. This method would work well in certain cases, such as undercover in programming advertisers that pay to have their products featured in TV shows and movies. I believe it would work exceptionally well on the kind of advertising Apple and Microsoft do by having characters prominently feature the backs of their laptops. In the case of billboards, the entire advertisement would not be eliminated but it would be a first step in fighting back against real life advertisements. However, even in that case much of the company information would be eliminated.

Methodology / Approach

II. Proposed Work
I will be using a three step method to come to the ultimate goal of removing the logo from the image.
In step one, I will build a convolutional neural network that creates bounding boxes around logos. I will be using the dataset collected from Flickr and annotated by the University of Ausburg for their research in logo recognition. This dataset will be trained on a convolutional neural network based on YOLO V2 and V3 by Joseph Redmon and Ali Farhadi to detect the logos and create the bounding boxes. The algorithm will be evaluated on its accuracy, precision, and recall of identifying logos in images.
Step 2 will consist of using a technique like SIFT to align clean images of the logos to the real world images. Clean images of the logo and correct alignment will be important in the final step.
In step 3, I will have mask of the logos that will be used for the final step, inpainting the logos. Ideally if time allows I would like to use one of the better inpainting algorithms, such as Exemplar based image inpainting by Criminisi et al.

Technologies Used

Python, Intel AI Dev Cloud