Deep learning/Image recognition- How to compose dataset for one category recognition

Question

Im using Intel Deep learning SDK (Based on Caffe).
I want to compose photos dataset for the ability of recognizing coins.
The SDK is expecting dataset in the following format:

Root
  --> category a
    --> photos
  --> category b
    --> photos
  ...  

And it required to have minimum 2 categories.
Any idea?

Answer 1

In your case you should have a "coins" category, and "non coins" category.
Your net should see images of "non coins" in order to distinguish these from "coin" images.

You might consider training in stages (aka "hard negative mining"):

(1) model_0 <- training with cons/non-coins images
(2) for k = 1..K iterations do:
(2.1)   use model_{k-1} to classify additional training images
(2.2)   add images wrongly classified by model_{k-1} to training set
(2.3)   model_k <- training with extended training set
(3) output model_K