Color segmentation and neural networks for automatic graphic relief of the state of conservation of artworks

This paper proposes a semi-automated methodology based on a sequence of analysis processes performed on multispectral images of artworks and aimed at the extraction of vector maps regarding their state of conservation. The graphic relief of the artwork represents the main instrument of communication and synthesis of information and data acquired on cultural heritage during restoration. Despite the widespread use of informatics tools, currently, these operations are still extremely subjective and require high execution times and costs. In some cases, manual execution is particularly complicated and almost impossible to carry out. The methodology proposed here allows supervised, partial automation of these procedures avoids approximations and drastically reduces the work times, as it makes a vector drawing by extracting the areas directly from the raster images. We propose a procedure for color segmentation based on principal/independent component analysis (PCA/ICA) and SOM neural networks and, as a case study, present the results obtained on a set of multispectral reproductions of a painting on canvas.