Home // IARIA Congress 2024, The 2024 IARIA Annual Congress on Frontiers in Science, Technology, Services, and Applications // View article

Prediction of Centroid Pixel Values in Image Triangulations Using a Graph Neural Network

Authors:
Luka Lukač
Andrej Nerat
Damjan Strnad
Filip Hácha
Borut Žalik

Keywords: image processing; Delaunay triangulation; machine learning; Graph Neural Network; interpolation.

Abstract:
Image triangulation is a simple and abstract representation of an image. Important image structures are represented with triangles that are scattered across the image in an unstructured manner. However, quite often, when dealing with an image triangulation, the user's object of interest is the original image. Various interpolation methods have been used in order to predict the original pixel values inside the triangulation simplices. Although yielding accurate results in some cases, their results can be significantly inaccurate when dealing with high-frequency details in simplices of the image triangulation, or if the triangulation simplices have a highly irregular structure. In this paper, a new interpolation method based on a graph neural network is proposed. The experimental results on the popular dataset DIV2K showed that the proposed method, in most cases, produces smaller prediction errors than the existing interpolation methods, such as Barycentric Coordinates or Inverse-Distance Weighting.

Pages: 106 to 110

Copyright: Copyright (c) IARIA, 2024

Publication date: June 30, 2024

Published in: conference

ISBN: 978-1-68558-180-0

Location: Porto, Portugal

Dates: from June 30, 2024 to July 4, 2024