Comparative analysis of convolutional neural networks and bilinear interpolation for mobile applications
Bulletin of Cherkasy State Technological University · 26.06.2026 · 31 (2) · 27-34
DOI: 10.62660/bcstu/2.2026.27 ↗Kurz gefasst
Fachartikel in der Zeitschrift Bulletin of Cherkasy State Technological University, online erschienen am 26.06.2026. Verfasst von Aleks Mosisa und 1 weiteren Autor:innen.
Abstract (Originalsprache)
The relevance of the study is conditioned by the growing need for efficient image processing algorithms for mobile devices with limited computing resources. The purpose of the study was to comprehensively compare the efficiency of convolutional neural networks and bilinear interpolation to predict the parameters of geometric transformations in mobile applications. The study was conducted on a Motorola G32 mobile device using the dual-input architecture of convolutional neural networks. A total of 120 experiments were conducted using images measuring 224×224 pixels, which included rotation angles ranging from -45° to +45° and scaling factors ranging from 0.5 to 1.5. The CNN model was optimised using TensorFlow Lite with 8-bit quantisation. The PSNR and SSIM metrics were used to evaluate image quality. The statistical analysis included: the Shapiro-Wilk test for normality verification, the Student’s t-test, and the Mann-Whitney U-test for group comparison, the Cohen coefficient for estimating effect size, ANOVA analysis of variance, Pearson and Spearman correlation analysis, and regression analysis. The thermal regime and energy consumption of the device were monitored at the significance level p < 0.05. Convolutional neural networks showed statistically significantly better image quality: the median PSNR value was 45.56 dB compared to 42.42 dB for bilinear interpolation, SSIM 0.953 and 0.910, respectively. Furthermore, bilinear interpolation provided perfect accuracy for predicting geometric parameters (median angle error 0.0°), in contrast to convolutional neural networks (median error 47.95°). The processing time for convolutional neural networks was 447.0 ms versus 358.9 ms for bilinear interpolation of battery consumption: 18.3 ± 4.2 mAh (CNN) versus 16.7 ± 3.9 mAh (bilinear), p > 0.05. The processor temperature increased by 3.2°C (CNN) versus 1.9°C (bilinear). Practical significance. The developed methods can be used in mobile augmented reality applications for optimal selection of algorithms depending on the application requirements
Zitieren
APA
Mosisa, A., & Vlasiuk, H. (2026). Comparative analysis of convolutional neural networks and bilinear interpolation for mobile applications. Bulletin of Cherkasy State Technological University, 31(2), 27-34. https://doi.org/10.62660/bcstu/2.2026.27
Vancouver
Mosisa A, Vlasiuk H. Comparative analysis of convolutional neural networks and bilinear interpolation for mobile applications. Bulletin of Cherkasy State Technological University. 2026;31(2):27-34. doi:10.62660/bcstu/2.2026.27
BibTeX
@article{Mosisa2026,
author = {Mosisa, Aleks and Vlasiuk, Hanna},
title = {Comparative analysis of convolutional neural networks and bilinear interpolation for mobile applications},
journal = {Bulletin of Cherkasy State Technological University},
year = {2026},
volume = {31},
number = {2},
pages = {27--34},
doi = {10.62660/bcstu/2.2026.27}
}
EUVIMED