Felippe Durán V. G. Santos, Renato Candido, Magno T. M. Silva. Avaliação de funções-custo na super-resolução de imagens para jogos eletrônicos. Simpósio Brasileiro de Telecomunicações – SBrT’2024, 2024, Belém, Anais do SBrT 2024, 2024, pp. 1–5. (in portuguese)

Abstract

Super-resolution image reconstruction has been of interest in several areas, such as medical images, surveillance, electronic games, among others. In this work, different cost functions are evaluated in the context of image super-resolution for electronic games. For this, a simplified version of the model of [1] is used, which uses multiple frames, their respective depth maps, and motion vectors. In addition to traditional cost functions such as the mean squared error, the mean absolute error, and the derived from the structural similarity index, we evaluate the the perceptual loss and the G-loss. Experimental results show that the choice of different perceptual loss layers influences the performance for the better, and the combination of those cost functions with the perceptual loss can result in details that improve the qualitative perception of the images. [1] L. Xiao et al., “Neural supersampling for real-time rendering,” ACM Trans. Graph., vol. 39, no. 4, 2020.

Keywords

Image super-resolution, deep learning, convolutional neural networks, video games, loss functions.

Downloads

• PDF