

Venation network analysis stands as a promising direction of scientific research, providing new insights into the origins and influence of plant phenotypic traits. However, its applicability is limited by a lack of tools that would facilitate relevant data acquisition. Dicotyledons form complex reticulate networks, often elusive for regular scanning equipment, hindering the attempts to capture details of their anastomoses arrangement. Currently available professional solutions operate on high-resolution noise-free images obtained in a complex process of chemical clearing, sample staining, and computationally expensive digitizing. This work introduces a novel technique capable of detecting leaf vasculature on pixel level and extracting a graph representation of its structure while operating on lower resolution scans of unprocessed specimens. The proposed transformation pipeline is designed as an array of steps — featuring automatic leaf segmentation, machine learning-based vein recognition, a sequence of custom spatial and morphological filters, segment radii retrieval and, finally, a graph compression and denoising algorithm. Each of those stages was separately evaluated using a range of metrics, including a new one aimed at assessing the uniformity of the reconstructed network. Obtained results confirmed that the method performs well in terms of both qualitative and quantitative analysis, given the characteristic imperfections in the examined images.