In this paper, we propose a method to estimate the concentration of pigments mixed in a painting, using the encoder-decoder model of neural networks. Encoder-decoder model is trained to output value which is same as input and its middle output extracts a certain feature as compressed information of the input. In this instance, the input and the output are spectral data of a painting. We trained the model to have pigments concentration as compressed information as a middle output. We used the dataset which was obtained from 19 pigments. The dataset has scattering coefficient and absorption coefficient of each pigment. We applied Kubelka-Munk theory to the coefficients to obtain many patterns of spectral data. It's shown that the accuracy of estimation is very high, and the speed of execution is very fast compared with a conventional method using simple for-loop optimization. We concluded our method is more effective and practical.
Kensuke Fukumoto, Norimichi Tsumura, Roy Berns, "Estimating concentrations of pigments using encoder-decoder type of neural network." in Proc. IS&T 27th Color and Imaging Conf., 2019, pp 149 - 152, https://doi.org/10.2352/issn.2169-2629.2019.27.28