A neural network to predict the spectral reflectance of prints

Abstract

The reconstruction of spectral reflectance from RGB triplets created by digital cameras is a topic of great interest. Different approaches dealing with this topic have been published. In the recent years, most approaches utilize neural networks. These approaches mainly differ in the chosen network architecture, the way of obtaining the dataset and the used hardware. While the most approaches aim for generalized applicability on a wide range of spectra, this paper aims for applicability on a limited set of spectra given by a typical use case of the printing industry. In this paper a neural network was trained to predict the spectral reflectance of prints. Therefore, 10 800 color patches were printed, measured by a spectrophotometer and captured by an RGB camera under different light sources generated with a DLP projector. The performance of the trained network was tested by determining the CIEDE2000 color difference as well as the mean squared error between the predicted and the measured spectral reflectance. The dataset was systematically reduced to examine how the number of color patches and light sources used for training influences the performance of a network. This paper shows that a network performed best when confronted with prints printed on the same substrate using the same color management settings as the dataset used for training. Training a network with multiple datasets on different substrates increased the generalization of a network, but decreased the performance compared to a network trained with a single combination of substrate and color management settings. Reducing the number of color patches as well as reducing the number of light sources influenced the performance of a network negatively, but still leads to decent results.