Neural Network-Based Digital Camera Spectrophotometer: Application to Chromogenic Technologies Characterization

Abstract

In this work, a neural-network-assisted procedure for transmittance spectral reconstruction from digital images is proposed. This allows the use of digital cameras as visible range spectrophotometers. The neural network is able to reconstruct a whole visible spectrum (in the range of 400 to 700 nm) from three-color coordinates. Precise color coordinates obtained from digital images are shown for different lighting conditions and file formats. Through a two-stage training of the neural network, consisting in a first broad gamut of transmittance spectral data set (3660 spectra) and a subsequent fine retraining data set (915 additional spectra), reconstruction of visible spectra with mean absolute errors below 2.15% (compared with measured spectra through a conventional spectrophotometer) is achieved. The possibility of simultaneous transmittance quantification over extended areas through digital imaging (statically or dynamically through video acquisition), which presently is restricted to a few mm 2 in conventional spectrophotometers, opens the way for improved colorimetry and visible range spectrophotometry. This could result in particular usefulness in the field of chromogenic technologies, where an adequate quantification of the homogeneity of color transitions over big surfaces is a key feature for a proper technological advance. For this purpose, the experimental application of the proposed system as a spectrophotometer was assessed through a complete optical characterization of four electrochromic polymers, obtaining the most relevant performance parameters (optical contrast, switching speed, and cycling stability). Excellent agreement between the values obtained from measured and reconstructed spectra proves the viability of the proposed tool, constituting a lowcost reliable alternative that overcomes some of the limitations of present conventional spectrophotometers.