A Bayesian method of generating color correction matrices for digital image sensors is presented. This method was developed for sensors with poor colorimetric quality, and uses strong prior assumptions about object surface reflectance functions to improve color correction accuracy. These assumptions are expressed through linear model constraints on surface reflectance coupled with a Normal distribution over linear model weights. Results obtained with simulations and real camera images are presented. The Bayesian method works well for highly non-colorimetric sensors and has been used in industrial practice with good and stable results. For sensors with better colorimetric properties, methods employing weaker assumptions about surfaces can sometimes produce better results.