Towards Semi-Automated Game Analytics: An Exploratory Study on Deep Learning-Based Image Classification of Characters in Auto Battler Games

With the growing interest in player profiles and game analytics, semi-automated annotation can provide data which can serve as a source of information for enhancing game strategies and the overall gaming experience. This study proposes a tooling-based workflow to tackle this subject while using the mini game Chicken Dinner in Heroes Charge as an example. It includes the implementation of a Deep Learning network in MATLAB, which runs the classification process for a semi-automated annotation. To accomplish this, several game play videos were recorded and split into individual frames. After a manual selection, these serve as training and testing images for the network. The outcomes of this research include a reproducible workflow that enables a partial annotation for auto-battler games, as well as a system that facilitates such annotation for the Chicken Dinner mini-game in Heroes Charge. The conclusion indicates that semi-automated annotation of digital game play is possible. This study shows an approach to semi-automated annotation and presents promising prospects for future research and applications.