Deep learning for game genre classification from game posters

Abstract

This thesis aims to classify the genres of video games using a deep learning based model. The models, which were developed based on the ResNet50V2 and InceptionV3 architectures in the Keras library, were trained with game capsule images taken from the Steam platform and genre labels determined by game publishers. In addition, an online survey was conducted in which a group of participants classified games according to their genres. The results were used to comparatively evaluate the ability of both human participants and deep learning models to infer genres from the image content. The survey participants performed this task with more success than the ResNet50V2 model. Among the genres analyzed, it was observed that the prediction scores of the "indie" and "action" genres were higher than the "RPG" and "strategy" genres. This difference was also observed when the survey results were compared with the performance of the model using ResNet50V2. While the predictions made by survey participants show less variation between genres, the low performance of the ResNet50V2 model can be attributed to the different proportions of genres in the dataset. Therefore, it is recommended that the study be repeated with a more extensive data set, especially for these species. The variation in the survey results may be due to the ambiguity in the definition of RPG and strategy games or, in other words, the broader definition of these genres. The discrepancy in the ResNet50V2 model could be caused by the problem's structure or dataset size. Consequently, we trained two more models with the same number of data for every genre and one genre for every game. Again the ResNet50V2-based model performed poorly across most genres, and changing the dataset and issue format had no considerable effect on performance. Nonetheless, the model constructed using InceptionV3 demonstrated noteworthy advancements, especially with an RPG success rate of 82\% and an overall accuracy of 52\%. These findings imply that InceptionV3 outperforms ResNet50V2 in this challenge, suggesting that it is a more effective design. When the games that the ResNet50V2 model and the survey participants performed well and poorly were divided into two groups, it was observed that action, adventure, indie and casual game genres were more prominent in the well-performed group. It was stated that the images in the poorly performed group usually consisted of only the game name written on a simple background and that the common opinion that these images were incomprehensible was a positive result for the "providing design feedback" use case of the model. Likewise, the occurrence of cases where the model and survey participants marked the same genres but did not match the game genres on the Steam platform indicates that the model can be used for the "genre validation" use case. A noteworthy point between this study and previous studies is that there is no fixed standard for determining game genres. To overcome this problem and to find hidden relationships between games, unsupervised classification studies can be carried out in the future, and genres can be grouped into clusters without naming them, and the consistency of these clusters can be examined. Based on the responses from survey participants, the game images were reclassified, and new models capable of generating genre-specific game images were created using the stable diffusion 1.0 model. Subsequently, these models were employed to generate sample images for imaginary games, and a new survey was conducted. In this survey, the same pool of participants was asked to match the images with the corresponding genres. Unlike the initial survey, the independent game genre, which defines the organizational structure of the game developer team, and the casual game genre, which describes the complexity of the game, were excluded from this study based on the feedback received. The results of the second survey were similar to those of the first survey. When compared to the base model, the genre-specific image generation model produced images with higher genre predictability. As a result, using a deep learning models in recommendation systems based on game genres will improve the performance of the recommendation system and genre-specific image generation model has potential to assist designers. This thesis study contributes to game development and marketing processes by revealing the power of game images to reflect genres and the difficulties in accurately determining genres.