Analysis of visual political communication on YouTube

Abstract: Though images are ubiquitous in everyday life and have always been part of politics, research on the visual aspects of political communication recently gained momentum, especially with the rise of social media. This opens up a platform to analyze the role of visuals in communicating political ideas. Images are a key part of the communication process, shaping peoples’ attitudes and policy preferences on political ideas. Generally iconic themes dominate representation of political ideas, for example iconic themes like polar bears represent the issues of climate change and environmental policies. This thesis focuses on finding such distant iconic themes in visuals of growing social media platforms like YouTube using deep learning. The initial analysis revealed the poor performance of the existing state-of-the-art networks on image classification in detecting the simple iconic theme of Polar bear in visuals. This arises a need for a new approach to improve the performance in detecting visual themes. The thesis proposes a method to develop a custom network by transferring the knowledge of the state-of-the-art networks using transfer learning. The result shows that the custom network has a better recall on predicting Polar bears than the state-of-the-art networks and the impact of training methods on predicting visualthemes on YouTube data.