In my opinion the peer-review should be changed to a double-blind system: the reviewer should not see name and affiliation of the authors, and judge the work as it would grade an undergrad paper (i.e. harshly). Like this I believe the signal-to-noise ratio in journals would increase, and only good papers would get published.
Please no! The problem with this approach (and it's already happening) is that what will get published is boring papers that bring tiny improvements over the state of the art. They'll get accepted because the reviewers will find nothing wrong with the paper, not because there's much good in there. On the other hand, the really new and interesting stuff will inevitably be less rigorous and probably more controversial, so it's going to be rejected.
Personally, I'd rather have 5% great papers among 95% of crap, than 100% papers that are neither great, nor crap, but just uninteresting. Reviews need to move towards positive rating (how many thing are interesting), away from negative ratings (how many issues you find in the paper). But it's not happening any time soon and it's one of the reasons I've mostly stopped reviewing (too often overruled by the associate editor to be worth my time).