Hierarchical clustering as an unsurpervised machine learning algorithm for hyperspectral image segmentation of films

In recent years, hyperspectral imaging has been increasingly applied to cultural heritage analysis, conservation, as well as digital restoration practice. However, the processing of such a large amount of data acquired by the instrument remains a challenging problem. In the absence of a machine learning pipeline, the segmentation task conducted in conventional methods is usually time-consuming and labor-intensive or requires pre-knowledge of the chemical nature of the art objects. In this paper, we propose to use the hierarchical clustering algorithm (HCA) as an alternative and necessary machine learning approach to segment the spectral images into clusters with different hierarchies, to maximize the information we can obtain from the high-dimensional hyperspectral data. Data were acquired using a custom-made push-broom VNIR hyperspectral camera (380-780 nm) and the materials used were a set of cinematic film samples with different degradation degrees. As preliminary results, different fading areas in the entire dataset are successfully classified and segmented. With the help of the automating and unsupervised algorithm, the effective segmentation of those degradation areas could be beneficial as it provides the basis for the future digital unfading treatment.