The question the chapter opens with
Zoom, enhance
Drag the box to choose a region, then throw away most of its pixels. Every method below gets exactly the same low-resolution evidence and has to produce a full-size picture from it. The leftmost tile is the truth none of them are allowed to see — so you can check their work.
Every pixel here is the average of a block of original pixels — which is what a sensor with larger photosites, or a video encoder, actually does. Averaging is not invertible: that is the whole problem.
This is the ground truth. In a real case it does not exist — that is the point of the exercise. Try the sign scene at 6× or 8× and see whether the characters survive.
PSNR is a pixel-difference score; SSIM is a structural one. They frequently disagree, which is the subject of a later demo. The outlined tile is the winner on SSIM among the four reconstructions.
So can you enhance it? Not in the sense the films mean. Interpolation — nearest, bilinear, bicubic — invents nothing at all; it only decides how smoothly to smear the measurements you already have. That is why none of those three tiles ever contains a detail the low-resolution input does not already imply.
Back-projection is the interesting one. It repeatedly asks "if my guess were the true image, what would the camera have measured?", compares that against the actual measurement, and pushes the difference back into the guess. It therefore guarantees something real: its output does downsample back to the evidence. But so do infinitely many other images, and nothing in the algorithm says which of them to prefer. That missing ingredient is the prior, and learning a prior from data is precisely what SRCNN and DLSS add.
Note what back-projection is not: it is not a neural network, and it has never seen another photograph in its life. It is included here as an honest stand-in, so that the gap between "consistent with the evidence" and "actually correct" is visible without a trained model in the page.