The Variogram as a Tool for Characterization The Variogram as a Tool for Characterization

Abstract

This paper explores the use of the experimental variogram as a descriptor for feature extraction in images. This characterization is evaluated using two classification models: a fully connected neural network (FCNN) and support vector machines (SVM). The MNIST dataset, which is widely used as a benchmark in image classification tasks, is employed for experimental validation. The study considers directional experimental variograms computed at orientations of 0°, 45°, and 90°, as well as the omnidirectional variogram as a baseline for comparison. In order to exploit directional information, two combination strategies are analyzed: a late fusion approach based on a weighted voting scheme applied to the class probability outputs of independent models, and an early fusion approach based on the concatenation of directional feature vectors. The results allow assessing the ability of the variogram to reduce the dimensionality of the problem while preserving spatial information that is relevant for classification.