Linear vs. Nonlinear Feature Combination for Saliency Computation: A Comparison with Human Vision