I've worked more on my dissertation project, I've got something coming together. Essentially images are taken from InTheBottle.co.uk, scaled down to a more reasonable size (still bloody huge) and then trained to the network the same way the colours were previously.
As you can see I think its working, it looks as though its clumping the images in the correct places but this could be just wishful thinking.
Again if anyone can assist me or has pointers let me know!