Reconstructing spectral reflectance from digital camera through samples selection

Spectral reflectance provides the most fundamental information of objects and is recognized as the "fingerprint" of them, since reflectance is independent of illumination and viewing conditions. However, reconstructing high-dimensional spectral reflectance from relatively low-dimensional camera outputs is an illposed problem and most of methods requaired camera's spectral responsivity. We propose a method to reconstruct spectral reflectance from digital camera outputs without prior knowledge of camera's spectral responsivity. This method respectively averages reflectances of selected subset from main training samples by prescribing a limit to tolerable color difference between the training samples and the camera outputs. Different tolerable color differences of training samples were investigated with Munsell chips under D65 light source. Experimental results show that the proposed method outperforms classic PI method in terms of multiple evaluation criteria between the actual and the reconstructed reflectances. Besides, the reconstructed spectral reflectances are between 0-1, which make them have actual physical meanings and better than traditional methods.