ЗГОРТКОВА НЕЙРОННА МЕРЕЖА З ПРОЕКТИВНО-IНВАРIАНТНИМ ПУЛIНГОМ

Abstract

This paper addresses the problem of image classification under projective transformations and proposes a convolutional neural network (CNN) architecture incorporating a projective invariant pooling layer. Unlike classical affine transformations, for which well-known equivariant transformations exist (e.g., steerable convolutional neural networks, harmonic H-Nets), the problem of finding projective equivariance remains open. This paper takes a step towards solving this problem and proposes an implementation of projectively invariant pooling. Compared to a standard CNN, we demonstrate that incorporating such pooling enhances the robustness of our network to projective distortions. Experiments are conducted on the proMNIST and rotoMNIST image datasets, generated from the standard MNIST dataset by applying corresponding transformations.