[ajtulloch]

Caffe (and other frameworks) provide exactly this. It's basically:

1) To setup, load a pre-trained AlexNet/Overfeat/other architecture model (e.g. trained on ILSVRC2012)

2) To get a vector from an image, run a forward pass on the images, and extract the activations at a given layer (e.g. fc7) as the output vector.

http://caffe.berkeleyvision.org/gathered/examples/feature_ex... is a step-by-step walkthrough.

There's a lot of mystique around deep learning and these kind of problems, but it's not _that_ difficult to use these tools.