Have you looked at PiCloud? It looks like exactly the base you want to build on. They run AWS instances, but charge by the millisecond. You can run one up just for the 5-15 minutes it takes.

Or if you want to get fancier, try their queues [0], they've got an example of doing image processing to explain how it works. Essentially, you set up your processing pipeline attached to a queue. You then drop messages on the queue whenever you want to do work, and they spin up one or many machines (controllable by you) to process the data. Then after 20 seconds of nothing new appearing on the queue, they turn them all off.

They'll even give you 20 hours of compute time free, every month.

Edit - if you don't want to build it, drop me a message, I can do it.

[0] http://blog.picloud.com/2013/04/03/introducing-queues-creati...