The Overfocusing Bias of Convolutional Neural Networks: A Saliency-Guided Regularization Approach