This could possibly be fixed by generating training image sequences which, along with random noise pixes, also have random hue shift for subregions of the image.