Study of Camera Spectral Reflectance Reconstruction Performance using CPU and GPU Artificial Neural Network Modelling

摘要

Reconstruction of reflectance spectra from camera RGB values is possible, if characteristics of the illumination source, optics and sensors are known. If not, additional information about these has to be somehow acquired. If alongside with pictures taken, RGB values of some colour patches with known reflectance spectra are obtained under the same illumination conditions, the reflectance reconstruction models can be created based on artificial neural networks (ANN). In Matlab, multilayer feedforward networks can be trained using different algorithms. In our study we hypothesized that the scaled conjugate gradient back propagation (BP) algorithm when executed on Graphics Processing Unit, is very fast, but in terms of convergence and performance, it does not match Levenberg-Marquardt algorithm (LM), which, on the other hand, executes only on CPU and is therefore much more time-consuming. We also presumed that there exists a correlation between the two algorithms and is manifested through a dependency of MSE to the number of hidden layer neurons, and therefore the faster BP algorithm could be used to narrow the search span with the LM algorithm to find the best ANN for reflectance reconstruction. The conducted experiment confirmed speed superiority of the BP algorithm but also confirmed better convergence and accuracy of reflectance reconstruction with the LM algorithm. The correlation of reflectance recovery results with ANNs modelled by both training algorithms was confirmed, and a strong correlation was found between the 3rd order polynomial approximation of the LM and BP algorithm's test performances for both mean and best performance.