Interpolation and masking effects of heteroassociative compressive transformations
M. A. Dryuchenko, А. А. Сирота · Journal of Physics Conference Series · 2021
Abstract The article considers heteroassociative image transformations, which perform mutual mapping of two neighbouring or nested regions of arbitrary shape within rectangular fragments or blocks covering the image. Specifically, the study focuses on heteroassociative compressive transformations. In the article, we provide theoretical justification of various types of heteroassocia-tive transformations performed using neural network based converters and numerical methods. We also study interpolation and masking effects of the transformations, which can be observed during image processing. These effects are accounted for by the possibility to represent an image as a sum of the deterministic component of the output (which is either linear of nonlinear function of the in-put) and the additive stochastic component of the output responsible for the prediction error. The article analyses occurrence, genetic and heuristic algorithms used to optimise the input and output configuration of the fragments covering the image. The algorithms used allowed us to maximise (minimise) the deterministic and the stochastic component and thus regulate interpolation and masking effects of heteroassociative transformations.