We have been using CouchDB for that, saving images as an “Attachment”. But after a year the multi-dozen GB CouchDB Database files turned out to be a headache. For example CouchDB replication still has issues if you use it with very large document sizes.
So we just rewrote our software to use CouchDB for image information and Amazon S3 for the actual image storage. The code is available at http://github.com/hudora/huImages
You might want to set up a Amazon S3 compatible Storage Service on-site for your project. This keeps you flexible and leaves the amazon option without requiring external services for now. Walruss seems to become the most popular and scalable S3 clone.
I also urge you to look into the Design of Livejournal with their excellent Open Source MogileFS and Perlbal offerings. This combination is probably the most Famous image serving setup.
Also the flickr Architecture can be an inspiration, although they don’t offer Open Source software to the public, like Livejournal does.