Image resizing by neural network operators and their convergence rate with respect to the $L^p$-norm and the dissimilarity index defined through the continuous SSIM

TL;DR

This paper develops new theoretical estimates for neural network-based image resizing operators, analyzing their convergence in $L^p$-norm and SSIM-based dissimilarity, supported by numerical experiments comparing with traditional methods.

Contribution

The paper introduces novel quantitative convergence estimates for neural network operators in image resizing, extending analysis to $L^p$-spaces and SSIM, with supporting numerical validation.

Findings

Neural network operators achieve high SSIM scores in image resizing.

Theoretical estimates align well with numerical results.

Proposed method outperforms bilinear, bicubic, and u-VPI in SSIM.

Abstract

In literature, several algorithms for imaging based on interpolation or approximation methods are available. The implementation of theoretical processes highlighted the necessity of providing theoretical frameworks for the convergence and error estimate analysis to support the experimental setups. In this paper, we establish new techniques for deriving quantitative estimates for the order of approximation for multivariate linear operators of the pointwise-type, with respect to the $L^p$-norm and to the so-called dissimilarity index defined through the continuous SSIM. In particular, we consider a family of approximation operators known as neural network (NN) operators, that have been widely studied in the last years in view of their connection with the theory of artificial neural networks. For these operators, we first establish sharp estimates in case of $C^1$ and piecewise (everywhere defined) $C^1$-functions. Then, the case of functions modeling digital images is considered, and specific quantitative estimates are achieved, including those with respect to the mentioned dissimilarity index. Moreover, the above analysis has also been extended to $L^p$-spaces, using a new constructive technique, in which the multivariate averaged modulus of smoothness has been employed. Finally, numerical experiments of image resizing have been given to support the theoretical results. The accuracy of the proposed algorithm has been evaluated through similarity indexes such as SSIM, likelihood index (S-index) and PSNR, and compared with other rescaling methods, including bilinear, bicubic, and upscaling-de la Vall\'ee-Poussin interpolation (u-VPI). Numerical simulations show the effectiveness of the proposed method for image processing tasks, particularly in terms of the aforementioned SSIM, and are consistent with the provided theoretical analysis.