Spectral reflectance reconstruction based on multi-target regression with two-directional stacking

A multi-target regression with two-directional stacking is proposed and applied to spectral reflectance reconstruction. Most of the previous studies for this problem have focused on proposing or improving the prediction model between RGB response space and spectral reflectance space, without considering the correlation between the components of the spectral reflectance vector. In this paper, we use the multi-target stacking regression for spectral reflectance reconstruction, the aim is to exploit the correlation between the output components. Furthermore, in order to overcome the overfitting or underfitting of traditional stacking regression method, we propose a novel two-directional stacking regression method. This paper also demonstrates that the stacking regression method has no effect with linear models and is only effective with nonlinear models. Experimental results show that the multi-target stacking regression is better than the no-stacking regression, and the multi-target regression with two-directional stacking is better than using the traditional one-directional stacking.