Computer vision is a powerful tool for intelligent sensor development. However, noise in CCD cameras leads to significant radiometric distortions. Therefore, radiometric image correction is a critical operation, specially when physics-based models are applied for image processing, as is the case in many industrial applications. All known radiometric correction methods assume that noise characteristics remain stable over time. In this paper, a new radiometric correction method is proposed to account for non static noise effects. The method decomposes radiometric distortion into multiplicative and additive errors, whose optimal models are computed with a new extension to the Generalized Cross-Validation Criterion.