<p>Classical image classification methods, including recent Graph Neural Network approaches, often rely on the direct use of raw images or implicitly learned graph structures, offering limited control over the explicit topological representation of image content. In contrast, this work is motivated by the idea of preparing a structured graph-based input for image classification through an explicit image-graph transformation. We propose a meshing-based framework that converts an image into a well-defined undirected graph, serving as a topological representation of the image. Two graph construction strategies are investigated: (i) a regular mesh based on minimum Euclidean distance and (ii) an irregular mesh derived from a Voronoi-based approach. The resulting adjacency matrices encode the structural organization of the image and constitute a suitable input for graph-based analysis. Unlike conventional GNN pipelines that operate directly on full images, the proposed approach introduces an intermediate and controllable graph construction stage, which enables the subsequent computation of topological indices. Once the image is transformed into a graph, a wide range of topological descriptors can be computed using existing graph analysis tools and software, providing a rich and interpretable feature space for image classification. The present study deliberately focuses on this fundamental graph generation stage, which forms the cornerstone of the proposed framework. The computation of topological descriptors and their integration into learning models are identified as natural perspectives, enabled by the graph representations established in this work.</p>