Scholarlink

Nowadays, video games have become widespread and popular. Game development is quite a complex process, often located at the “cutting edge of technical complexity”. Like any other large project, it is accompanied by technical problems. Graphic rendering errors not only spoil the gaming experience of players, but also cause reputational and financial damage to developers and publishers. In this paper, we propose an approach based on Image-Feature-Informed Networks (IFINet) for anomaly detection (AD) in video game frames. It consists of using two ideas: (1) we try to find meaningful image representations (feature vectors), (2) use shallow AI models on the extracted features. Such representations include histograms, Discrete Fourier transform (DFT), Sobel gradients, etc., which are then fed to the input of a neural network. This approach allows us to achieve fast model training and excellent quality metrics. In most cases, we have achieved TPR above \(95\%\) and FPR below \(1\%\) , which is better than current global rates.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Image-Feature-Informed Networks for Anomaly Detection in Video Game Frames

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Alexander Trykin,
  • Nikolai Sokolov,
  • Konstantin Pukhkii,
  • Evgenii Vasiliev,
  • Alexandra Serebriakova,
  • Vadim Turlapov

摘要

Nowadays, video games have become widespread and popular. Game development is quite a complex process, often located at the “cutting edge of technical complexity”. Like any other large project, it is accompanied by technical problems. Graphic rendering errors not only spoil the gaming experience of players, but also cause reputational and financial damage to developers and publishers. In this paper, we propose an approach based on Image-Feature-Informed Networks (IFINet) for anomaly detection (AD) in video game frames. It consists of using two ideas: (1) we try to find meaningful image representations (feature vectors), (2) use shallow AI models on the extracted features. Such representations include histograms, Discrete Fourier transform (DFT), Sobel gradients, etc., which are then fed to the input of a neural network. This approach allows us to achieve fast model training and excellent quality metrics. In most cases, we have achieved TPR above \(95\%\) and FPR below \(1\%\) , which is better than current global rates.