Graham D. Finlayson

Colour correction is the process of converting RAW RGB pixel values of digital cameras to a standard colour space such as CIE XYZ. A range of regression methods including linear, polynomial and root-polynomial least-squares have been deployed. However, in recent years, various neural network (NN) models have also started to appear in the literature as an alternative to classical methods. In the first part of this paper, a leading neural network approach is compared and contrasted with regression methods. We find that, although the neural network model supports improved colour correction compared with simple least-squares regression, it performs less well than the more advanced root-polynomial regression. Moreover, the relative improvement afforded by NNs, compared to linear least-squares, is diminished when the regression methods are adapted to minimise a perceptual colour error. Problematically, unlike linear and root-polynomial regressions, the NN approach is tied to a fixed exposure (and when exposure changes, the afforded colour correction can be quite poor). We explore two solutions that make NNs more exposure-invariant. First, we use data augmentation to train the NN for a range of typical exposures and second, we propose a new NN architecture which, by construction, is exposure-invariant. Finally, we look into how the performance of these algorithms is influenced when models are trained and tested on different datasets. As expected, the performance of all methods drops when tested with completely different datasets. However, we noticed that the regression methods still outperform the NNs in terms of colour correction, even though the relative performance of the regression methods does change based on the train and test datasets.

Recently, many deep neural networks (DNN) have been proposed to solve the spectral reconstruction (SR) problem: recovering spectra from RGB measurements. Most DNNs seek to learn the relationship between an RGB viewed in a given spatial context and its corresponding spectra. Significantly, it is argued that the same RGB can map to different spectra depending on the context with respect to which it is seen and, more generally, that accounting for spatial context leads to improved SR. However, as it stands, DNN performance is only slightly better than the much simpler pixel-based methods where spatial context is not used. In this paper, we present a new pixel-based algorithm called A++ (an extension of the A+ sparse coding algorithm). In A+, RGBs are clustered, and within each cluster, a designated linear SR map is trained to recover spectra. In A++, we cluster the spectra instead in an attempt to ensure neighboring spectra (i.e., spectra in the same cluster) are recovered by the same SR map. A polynomial regression framework is developed to estimate the spectral neighborhoods given only the RGB values in testing, which in turn determines which mapping should be used to map each testing RGB to its reconstructed spectrum. Compared to the leading DNNs, not only does A++ deliver the best results, it is parameterized by orders of magnitude fewer parameters and has a significantly faster implementation. Moreover, in contradistinction to some DNN methods, A++ uses pixel-based processing, which is robust to image manipulations that alter the spatial context (e.g., blurring and rotations). Our demonstration on the scene relighting application also shows that, while SR methods, in general, provide more accurate relighting results compared to the traditional diagonal matrix correction, A++ provides superior color accuracy and robustness compared to the top DNN methods.