It looks like a tough subject. Finding a suitable feature space to make your problem solvable is probably going to take some hard work and experimenting. Anyway, challenges are always good! :)
If you provide your background it would help a lot answering your question. Assuming you're new to the subject:
Take a look at libsvm, read the papers, run the example and learn the python bindings. You can also approach the problem using neural networks or other classifiers but at least between my peers that work in research (not many, I must admit) SVM seems more common today.
Since you're not going to do pixel to pixel classification - only the whole image matters, not individual pixels - you can have a rough start by classifying histograms from different color spaces (histograms for R, G, B, H, S, V, mean values, standard deviations). I'd try a Hough transform and other feature extraction techniques to expand the feature space into something broader.
Good luck :)