Speaker
Description
Spectral imaging techniques, such as Macro Area X-RayFluorescence (MA-XRF), are extremely powerful methods to investigate the composition of pictorial artworks. The produced data contains a vast amount of information about the composition of the pictorial artwork under scrutiny; nevertheless, post-processing analyses to extract such meaningful information are usually complex, lengthy, and require days (if not weeks) of work by experienced heritage scientists.
A possible approach to tackle these issues is to apply Machine Learning methods in order to extract meaningful information from spectral imaging raw data. In this contribution, we explore the use of unsupervised methods for a fast analysis of MA-XRF raw data. The idea is to see the single-image MA-XRF datacube, i.e. a tensor of shape (𝐻,𝑊, 𝐷),where 𝐻 is the number of pixels in the image height, 𝑊 is the number of pixels in the image width, and 𝐷 is the number of energy bins, as a dataset of spectra of shape (𝑁 = 𝐻 · 𝑊, 𝐷).On this dataset, we perform Principal Component Analysis (PCA), to extract a lower-dimensional representation of the spectra, and then we apply the fuzzy Gustafson-Kessel clustering algorithm to assign to each sample point (i.e .each pixel) a weight, which loosely represent the probability that each pixel belongs to a certain cluster. To speed up the algorithm, we used the JAX framework to move the computation from CPU to GPU, reducing by at least an order of magnitude the execution time (from minutes to seconds).
By rearranging the weight matrix into a set of single-channel images, we get a grayscale image, we can extract visual information in the form of (soft) semantic segmentation. The fuzzyness of the clustering, i.e. the fact that we have grayscale images and not binary maps out of the clustering process, gives additional visual feedbacks to heritage scientists, helping also to identify anomalous pixels in the images, which may present interesting features