CSI zoom-in is clearly bogus, for good, information-theoretic reasons. But there is a grain of truth in CSI blur removal: where you've got a good model for the blur, e.g. motion blur or Gaussian (camera-out-of-focus) blur, maximum entropy modelling can do a startlingly good job of computing the original image most likely to have produced the blurred image, though it's very computationally expensive. (And, just as with Mark's point above, you can't do it on JPEGs, which have already thrown away as "spurious" the fine detail that it depends on.)

Peter