| Abstract [eng] |
The increasing growth in video games industry raises the need for automated content curation and age rating. No solutions for this problem were found in the literature. In this work we attempt to solve the problem using deep neural networks. A data set for this problem was created. It contains over a thousand different video games, 11 thousand video excerpts, text descriptions, PEGI and ESRB ratings and other meta data. Based on the literature solving similar classification problems (violence detection in videos, gunshot sound detection in sound, etc.), we created more than 12 neural networks. Networks showing the best accuracy were further composed using different methods, thus making classification consider all available data – video frames, optical flows, sounds and text descriptions. Composed networks showed better accuracy compared to individual ones: one was good at classifying children (93 %) and adult (81 %) games, the other had the highest overall accuracy (63 %). The problem was found to be difficult. The most difficult part was to classify games rated for teenagers, that is due to how similar they are either to children or adult games. The other difficult part was the abundance of data and low saturation of relevant information: videos are long, but only a part of them contains important information. Text descriptions proved to be superior compared to other data types in this problem, due to their high density of important information they showed the best accuracy among individual data type networks. |
