The colour of objects is most often described using images captured with a commercial RGB camera. Such a description depends on the characteristics of the capturing device and the object illumination. Independent of these factors is the object description using the reflectance spectrum, which can be captured by a spectrophotometer for point readings and by a multi- or hyper-spectral camera for larger objects. The latter are expensive, which stimulates research into the possibility of mapping RGB camera images to the reflectance spectrum. Many methods have been proposed, ranging from purely mathematical models to artificial neural network (ANN) modelling approaches of different architectures and complexity. Most modelling approaches need information on camera properties and illumination. In the ANN method presented in our study, which is based on a simple, fully connected neural network with nonlinear activation functions of neurons in hidden layers and a smaller number of inputs and a larger number of outputs, the modelling does not require knowledge of the camera characteristics, its sensors and illumination, as the data for ANN learning is obtained by simultaneously capturing the object and the reference samples. Special attention was paid to determine the influence of hyperparameters on the performance of reflectance spectrum reconstruction using ANN models with respect to the choice of the learning algorithm, the size of the training set, the number of input data – RGB readings or the number of cameras, the number of neurons in the hidden layers, and the number of hidden layers, setting five initial working hypotheses and investigating their validity in a three-step study, described in SCI-indexed journal articles. The probability of finding successful ANN models increases with more modelling iterations for the chosen configuration, but finding more successful models is time-consuming. Two approaches are proposed to make the search for ANN models more efficient. The first procedure, using a faster but less efficient single-layer ANN learning algorithm executed on a graphics processor, narrows the search area for a second, slower, more efficient algorithm executed on a central processing unit, and the second procedure, which, given a selection of hyperparameters of the ANN models and a selection of a criterion function, proposes a search centre as the number of neurons in the hidden layers, around which to search for the best-performing ANN models in a narrower neighbourhood.
|
