Game Genre Classification from Icon and Screenshot Images Using Convolutional Neural Networks
Chayanin Suatap, Karn Patanukhom · 2019
In this paper, we study game genre classification using typical game images provided in game stores such as an icon or screenshots. The proposed method is based on a convolutional neural network and a soft voting ensemble technique. Network depth and dropout strategy are examined to obtain the best network architecture for performing the task. The ensemble technique is applied to boost the classification accuracy. Experiments are conducted on Android game dataset that consists of 25,001 icon images and 180,553 screenshots from 17 game genres. Our proposed method can achieve 40.3% and 46.7% classification accuracies for single icon and screenshot classification tasks, respectively. It also outperforms the human performance in both tasks. In addition, it can provide accuracy of 55.3% when multiple screenshots and icon of the game are used together for classifying their genre.