Technical gazette, Vol. 29 No. 3, 2022.
Original scientific paper
https://doi.org/10.17559/TV-20210405174515
Exploiting Nonlinearity between Parallel Channels of Multiple Cameras for Accurate ANN Reconstruction of Reflectance Spectra
Mihael Lazar
; University of Ljubljana, Slovenia, Faculty of Natural Sciences and Engineering, Department of Textiles, Graphic Arts and Design, Snežniška ulica 5, SI-1000 Ljubljana, Slovenia
Aleš Hladnik
; University of Ljubljana, Slovenia, Faculty of Natural Sciences and Engineering, Department of Textiles, Graphic Arts and Design, Snežniška ulica 5, SI-1000 Ljubljana, Slovenia
Abstract
Colour of an observed object is unambiguously described by its reflectance. Translation from a colour description in RGB space obtained with a digital camera into reflectance, independent of illuminant and camera's sensor characteristics, was performed through an artificial neural network (ANN). In the study, it was hypothesized that the ANN's performance of reflectance reconstruction could be improved by using extended learning datasets with two or three cameras RGB input sets instead of one, but only if the parallel channels of cameras used are not linearly dependent. Nonlinearity was assessed by a quantitative measure of nonlinearity (QMoN), the metric primarily developed for use in chemistry. A noticeable reflectance performance improvement has been found with two and three cameras, even though the cameras' parallel channels exerted only small degrees of nonlinearity. Close attention was paid to the impact of scattering of RGB readings around the ideal values on the reflectance reconstruction performance, and it has been found that the more pronounced scattering is inversely proportional to the performance of ANNs trained with a single-camera input learning set but shows no visible impact on the performance of ANNs trained with extended input learning sets.
Keywords
artificial neural network; extended learning set; multiple cameras; parallel channel's nonlinearity; reflectance spectra reconstruction
Hrčak ID:
275280
URI
Publication date:
19.4.2022.
Visits: 951 *