A Novel White-Balance System for Broadcast Cameras Using Machine Learning

Color constancy is the ability to perceive colors of objects, invariant to the color of the light source. Our eyes, the human visual system is capable of doing that. Color constancy algorithms first estimate the color of the illuminant light source and then correct the image so that the corrected image appears to be taken under a canonical light source. In this project, we use color constancy on a broadcast camera without processing the video but instead acting on the RGB gain control much akin to a camera operator. The reference color is provided by a neural network which is trained to infer the color of the illuminant just looking at a sample frame from the camera; from that reference, we can compute the direction of the error from the white illuminant in color space and then generate a new RGB gain triplet to the camera. With the new adjustment, the cycle continues and eventually, the camera will output an image that is corrected and the neural network will estimate the illuminant as white yielding a zero error and the whole system stabilizes.

Edmundo Hoyle | GRUPO GLOBO | Rio de Janeiro, Brazil
Alvaro Antelo | GRUPO GLOBO | Rio de Janeiro, Brazil
Leandro Pena | GRUPO GLOBO | Rio de Janeiro, Brazil