In the Watson Image Recognition (IR) demo, you can give the app an image, and it will classify it, and tell you what it thinks is in the image. How does it achieve this?
So far as I can tell from the documentation, you can only give it positive and negative images, and it can only say 'Yes', this is Pizza, or 'No', this is not Pizza. It cannot specify whether it's pizza, or pasta.
Is there a way I can achieve this?
The service works by using a collection of classifiers; each classifier is for a single tag only and must be trained with it's own sets of positive and negative images. To get pizza vs pasta, you must create two classifiers and train each one with positive and negative images. (However, in that particular case, the negative images for pizza could just be the positive ones from pasta and vice versa.)
There is a default collection of classifiers available that has a fairly wide range of tags, all of which have been trained in this manner; I believe the demo uses a combination of that default collection + AlchemyVision (which has a wider range of classifications, but is not trainable). You can see the demo's source code on github.
You can read a bit more in the Visual Recognition service documentation: http://www.ibm.com/smarterplanet/us/en/ibmwatson/developercloud/doc/visual-recognition/index.shtml
