Spectral reflectance estimation based on two-step k-nearest neighbors locally weighted linear regression

Abstract

To improve the estimation accuracy of spectral reflectance from the given trichromatic value, a new two-step k-nearest neighbors locally weighted linear regression method is proposed. The algorithm has good local learning ability and can take into account the similarity of colorimetric and spectral reflectance space. The simulated and practical imaging experiments were carried out with Munsell matte and glossy dataset, respectively. Experimental results show that the mean root mean square error values of the spectral reflectance estimated by our model in simulated RGB, practical imaging Adobe RGB. and raw RGB data experiments are 0.00731, 0.01519, and 0.01453, respectively, and the mean color difference values under CIE standard illuminant D65 are 0.380, 1.311, and 1.180, respectively. In addition, we showed the calculation time cost of various models in the practical experiment. The calculation time of one sample for the proposed method is 0.094 s. The proposed method is better than several state-of-the-art methods in terms of comprehensive estimation performance and running efficiency.