A Model for Learning Variance Components of Natural Images