Implementation of a Neural Network for Multispectral Luminescence Imaging of Lake Pigment Paints

Abstract

Luminescence multispectral imaging is a developing and promising technique in the fields of conservation science and cultural heritage studies. In this article, we present a new methodology for recording the spatially resolved luminescence properties of objects. This methodology relies on the development of a lab-made multispectral camera setup optimized to collect low-yield luminescence images. In addition to a classic data preprocessing procedure to reduce noise on the data, we present an innovative method, based on a neural network algorithm, that allows us to obtain radiometrically calibrated luminescence spectra with increased spectral resolution from the low-spectral resolution acquisitions. After preliminary corrections, a neural network is trained using the 15-band multispectral luminescence acquisitions and corresponding spot spectroscopy luminescence data. This neural network is then used to retrieve a megapixel multispectral cube between 460 and 710 nm with a 5 nm resolution from a low-spectral-resolution multispectral acquisition. The resulting data are independent from the detection chain of the imaging system (filter transmittance, spectral sensitivity of the lens and optics, etc.). As a result, the image cube provides radiometrically calibrated emission spectra with increased spectral resolution. For each pixel, we can thus retrieve a spectrum comparable to those obtained with conventional luminescence spectroscopy. We apply this method to a panel of lake pigment paints and discuss the pertinence and perspectives of this new approach.