The method does have to identify lines of text to find the rotation angle, but doing so after perspective correction using the "all letters should be about the

The method does have to identify lines of text to find the rotation angle, but doing so after perspective correction using the "all letters should be about the same size" assumption means that a Hough transform is enough for that step, since the lines should already be roughly parallel. (Having to identify page boundaries is handwaved away with "I’m going to make a huge simplifying assumption that that the image we’re processing basically contains only the text that we want to rectify")