AdSegNet: a deep network to localize billboard in outdoor scenes

Abstract

Over the past decade, there has been a remarkable expansion in the popularity of online videos, due to the variety of content that has become accessible on the internet. This expansion presents an opportunity for advertising and marketing agencies to take advantage of targeted advertisements. Targeted advertisements can be accomplished by replacing an existing advertisement within an image frame with a new one. There is a limited amount of research on the general task of localizing billboard or advertisement board in outdoor scenes. Therefore, in this study, we proposed a deep neural network that uses a fusion of VGG16 and SegNet architecture to accurately identify the location of an advertisement in an image frame. To evaluate the effectiveness of our proposed method, we compare our proposed method to other semantic segmentation algorithms using a publicly available dataset of outdoor scenes annotated with binary maps of billboards. Our experimental results show that the proposed method achieves 98.58% training accuracy for billboard localization, while testing results gave 96.43% testing accuracy. Additionally, the low RMSE score of our AdSegNet model suggests that it can accurately determine the four corners of the billboard. Therefore, our approach could be beneficial to advertising and marketing agencies that seek to utilize targeted advertisements