Saliency-based selection of visual content for deep convolutional neural networks