A Convolutional Neural Network as a Proxy for the XRF Approximation of the Chemical Composition of Archaeological Artefacts in the Presence of Inter-microscope Variability.

The paper puts forward a convolutional neural network model for multi-output regression, which is trained on images from two distinct microscope types to estimate the concentration of a pair of chemical elements from the surface of archaeological metal objects. The target is to simulate the approximation behaviour of the more complex XRF technology, which is used as ground truth in training the model. Experiments investigate the adequacy of learning on either type of data and then using the models to test images coming from each microscope in turn and in combination. Under these terms of performance and flexibility, the technology can be successfully used in the front line of ancient object restoration across laboratories irrespective of the equipment available.