UCLA

Images are central to understanding protests and mass activism today for its impact in shaping public opinion.

Previously, analyzing protest images required human annotation, which is laborious and expensive. In the modern era of social media, an automated and systematic method is required to analyze the vast amounts of social media images.

In this thesis, I introduce a deep-learning computational framework to analyze protest images. This system comprises of (1) processing and parsing social media images from Twitter, (2) a model to identify common protest image characteristics, such as violence, fire, and police, models to (3) detect and (4) classify faces of protesters to understand demographics, and (5) an deduplication algorithm to identify the most shared images.