Detecting Visual Motives Using Online Clustering of Similar Images (Bachelor Thesis, Finished)
The proliferation of social media raises the question of how much and to what extend political opinions are formed and influenced by social media content. Recent studies have hinted that social media should not be underestimated in terms of its influence on elections. The goal of this work is to detect political themes, represented by clusters of images which were posted from different sources on social media. As a use case the 2019 Swiss federal election was chosen, Twitter data was gathered previously and processed to simulate an online real time data stream. The data stream is run through a processing pipeline to first contain only tweets with images. Since images on social media often undergo content modification or distortions, the images are then mapped to a perceptual hash, which enables detection of similar images. From the hash stream, clusters of similar images are created and finally the clusters are presented in a web application, where they are visualized based on two different approaches. First as a graph of clusters as nodes and similarity shown through edges and second as a bubble chart, where similarity is projected onto a two dimensional plane. The built system is evaluated based on it’s real life application feasibility, focusing on the processing time and user feedback. The results of the evaluation open up possibilities for further work, especially regarding different clustering methods and user interface extensions.
Start / End Dates
2020/02/24 - 2020/08/23