Training on new images

[Peter Lombrozo]

Sorry for being so new to this, but all I've seen so far is creating models based on image data already trained such as GoogleNet and Alexnet.  My question is can I use caffe for image recognition on unique, proprietary images?  I'd like to take a image set of jpgs, convert to lmdb, train on what in the image I'm interested in and use it on future input images.  How do I do this?  How do I label images to describe the part I'm interested in?  Its much like I have a set of squiggles on a page, but I'm only want to know if a particular squiggle is on that page.  How do I get started?  

I've done something similar using the Cognex system, but there, you mark the pattern you want and it matches the exact pattern in other images.  I'd like to use CUDA to do that instead, but recognize similar patterns, not exact matches.

Thank you for any help