Spectral reflectance estimation from non-raw color images with nonlinearity correction

Abstract

The spectral reflectance is recognized as the fingerprint of an object surface and has been used to achieve accurate color measurement in textile and other fields. Spectral reflectance can be recovered from color images to preserve high spectral and spatial resolutions simultaneously. However, a color camera commonly supplies a non-raw color image, which is non-linear with respect to the scene radiance, and is inappropriate for quantitative analysis. In this study, for non-raw color images, different nonlinearity correction models are designed and evaluated with respect to different spectral estimation algorithms. The colorimetric and spectral accuracy of spectral estimation after the nonlinearity correction is assessed through both simulation and practical experiments. In the simulation, a large number of spectral images from several datasets are employed to directly verify the effectiveness of the nonlinearity correction. In the practical experiments, the spectral estimation accuracy following the nonlinearity correction is verified directly and indirectly based on actual color images. The resulting linear color image data after the nonlinearity correction can provide better spectral estimation accuracy especially for the PI algorithm with one power-function based model. Besides, the combination of the simple PI algorithm with the power-function based model can exceed other combinations comprising complex algorithms and models in both accuracy and efficiency. For the linear color image data, the PI algorithm even surpasses the deep learning-based methods in certain metric, thus indicating a shallow relationship exists between the linear color image data and the spectral reflectance.