Dear Daniel,

Thank you a lot and instead of using PredictFuntion, we used classifier. This led to no significant improvement, as attached. On a second reading of your advice, it now looks to me that we should classify our images using the Classifier in the first place and then ensure that we have to use a large number of images for each class using PredictFunction. Cheers,

Attachments:

With roughly one or two images in each class for training, I'd be very surprised in a classifier could do much better than random in terms of predictions on a test set. Also it might be better to use Classify, if the intent is for the outcomes to be a discrete set that just happens to be integers.