Classifying Light States in the Phasmophobia Game. The Impact of Image Resolution and Preprocessing on Multi-layer Perceptron Performance

The paper explores the use of image preprocessing techniques to improve the classification of the number of lights turned on in the game Phasmophobia based on screenshots. A novel, publicly available dataset was created for this purpose, consisting of images captured under varied in-game conditions. The study examines the effects of color schemes, image sizes, normalization methods, and batch sizes on classification performance. Multiple preprocessing combinations were evaluated using a multi-layer perceptron model. Results indicate that the choice of color schema, normalization technique, and image resolution significantly influences classification accuracy. In particular, combinations such as subtractive color model used in printing, color separation (CMYK) with standard normalization at 64×64 resolution yielded the highest performance. Lower resolutions yield better results on average, suggesting that higher detail does not necessarily improve classification. Among normalization techniques, robust scaling and z-score normalization achieve the highest average accuracy, while transformations like discrete cosine transform and fast fourier transform perform significantly worse. These findings, based on averaged results over hundreds of configurations, highlight the impact of preprocessing choices on model performance.