What's the reasoning behind first scaling images to 256x256 and then cropping them to 227x227?

[Mario Klingemann]

If I am understanding the "finetune_flickr_style" solver.prototxt correctly all the images are first squeezed to 256x256 and then cropped to 227x227 since that's the data layers input size.

layer {
	name: "data"
	type: "ImageData"
	top: "data"
	top: "label"
	include {
	phase: TRAIN
	}
	transform_param {
	mirror: true
	crop_size: 227
	mean_file: "data/ilsvrc12/imagenet_mean.binaryproto"
	}
	image_data_param {
	source: "data/flickr_style/train.txt"
	batch_size: 50
	new_height: 256
	new_width: 256
	}
	}

 But looking at the caffe code that does the actual cropping it seems that the crop area is always a centered square. So this means that are border of about 15 pixels gets discarded all the time. I would understand it if the crop would always pick a randomly positioned square inside the image area, but since the result is the same all the time, why not scale the image to 227x277 from the beginning?