General advice about where to start on developing image recognition app

[Richard Rice]

First some background. I

collect weed species data in the field for the Washington state dept

of natural resources(DNR). There are several Android apps for

identifying weeds by looking at photos but none that I can find by

comparing a photo that I take with my Android phone with a

downloadable database of known weed species. What I'd like to do is

write an app that downloads DNR's weed photos of which there are about

1000 to my phone, find a weed in the field, take a photo and compare

it with DNR's photos to see if there is a match.

Does this sound feasible? I want to keep it simple, not try to

identify every weed in existence, just here in Washington.

More questions:

There are about 100 species of weeds. Each species has about 10 photos labeled with scientific,common names in various stages of development.

Should these 1,000 distinct photos be processed as multi-label or multi-class?

Would an SVM or Multi-classSVM be better than say a CNN algorithm? Maybe a nearest neighbor approach?

Could I make use of the Caffe framework?

Best Regards,

Richard