ABSTRACT This study presents a deep learning–based framework for modeling angular color variations on three‐dimensional surfaces under a standard lightbox using low‐cost RGB‐D imaging. A multilayer perceptron (MLP) model is employed to address multi‐angle color inconsistency by integrating nine‐dimensional feature information, including RGB sensor responses, spatial coordinates and surface normal to infer the CIE L * a * b * values of the color samples under nearly 0°:45° measurement geometry. The proposed method effectively captures high‐dimensional, nonlinear reflection behavior that cannot be fully predicted by traditional polynomial regression (PR) models. Experimental results demonstrate that the MLP forward model achieves an average color difference of 0.81 , significantly outperforming the PR model, which exhibits an average color difference of 6.39 . The inverse model, which predicts object color variations under the lightbox by using 0°:45° colorimetric data as input, achieves an average color difference of 1.02 . The proposed framework demonstrates robust performance under extreme measuring geometries and shows strong potential for applications that require simultaneous and accurate color measurement across multiple local regions of a 3D object.
Hung et al. (Mon,) studied this question.