With tremendous growth over the last decade [1], social media have become a central aspect of everyday life. Instagram is one of those fast-growing platforms focusing mainly on visual content. The shift from traditional media to online social media accompanying this growth does not stop at political campaigns [2].Therefore, the question arises to what extent visual personalization as part of political personalization shapes political communication on Instagram.
Political personalization describes the process of individual politicians playing an increasingly central role in political campaigns while political groupings are declining in power [3]. The concept of personalization as part of political communication already received plenty of attention. Numerous studies on this topic have been conducted across different countries and media types before the rise of social media. However, van Aelst et al. (2012) could not find conclusive evidence due to the absence of a common underlying concept or operationalization. To overcome this limitation, they suggested a concept to provide comparative results among international studies [4]. One of the more recent studies focusing on communication strategies on Instagram applied the conceptual idea from van Aelst et al. (2012) but still leaves room for improvement. One limitation of this work is the missing examination of Instagram stories due to their data collection process [5]. Furthermore, there is a lack of automatization of visual content analysis since they still rely on human labelers.
This lack establishes the need for automated visual content analysis methods, such as face recognition, in the context of political science. Despite various pre-trained models available, one issue remains unsolved: These models tend to suffer from a high false positive rate. To overcome this issue, methods to improve precision while maintaining a proper recall rate must be identified.
This work aims to achieve an improved face recognition performance for recognizing politicians in Instagram stories. More specifically, the goal is to reduce the false positive rate by integrating a face image quality assessment [6] step into the face recognition process. Therefore, a dataset of roughly 2200 images was acquired by scraping 21 parties' and their front runners' accounts. Those collected images are going to be hand labeled under consideration of an inter-annotator agreement. Subsequently, the effect of MagFace [8], a state-of-the-art face image quality assessment algorithm [6, 7], on the false positive rate is evaluated on the annotated dataset. Furthermore, the correlation between important people in the image and their suitability for face recognition is examined.