Prediction of RGB camera values by means of artificial neural networks

Abstract

Artificial neural networks (ANN) enable modelling of complex nonlinear systems that cannot be easily described using formal equations and have been implemented in many fields of science and technology for pattern recognition, clustering or data fitting. The goal of our study was to create a system that transforms XYZ and L*a*b* values into arbitrary camera RGB values in stable — but without strict knowledge of — photographing conditions, by means of the ANN data fitting ability. We adopted a two layer feed-forward neural network with sigmoid hidden and linear output neurons, that can fit multi-dimensional mapping problems quite well, when using enough neurons in the hidden layer and being fed by congruent learning set of data. The network was trained with Levenberg– Marquardt backpropagation algorithm. Learning data sets consisted of input XYZ or L*a*b* values and output RGB values. Input data were calculated from the reflectance values of Gretag Macbeth Digital ColorChecker SGtest chart obtained by spectrophotometric measurements, by taking into account three different standard illuminants (A, D50 and D65) and two standard colorimetric observers (2 ° and 10 °). Output data were RGB values of test chart ColorChecker SG acquired by Nikon D50 digital camera. Our goal was to find answers to several questions, such as what is an optimal number of hidden layer neurons, what degree of accuracy can we obtain by training ANN with a limited number of color samples, how does number of neurons affect ANN learning time and also which type of input data is more suitable for the prediction of RGB values. Since each ANN learning epoch starts with a random weight distribution and random training, validation and testing data selection, every learning cycle stopped in its local minimum. To assess the representative values of difference between the actual and the predicted values, learning cycle for each number of hidden layer neurons and for each learning data set was repeated many times and average ANN training time and average, median and minimal error rates for training, validation and testing data were recorded.