Doing this as a computer vision task would be very difficult for someone with little computer vision experience - 10 years ago it was an entirely unsolved problem. But to get you started:
Neural networks (or properly, NN with back-propagation-style training) are rather old hat, and no longer the method of choice. Random forests are popular, mostly because they quite flexible, reasonably easy to implement, and have on-average no worse performance that the other classification methods around. Criminisi et al 2011 is the standard paper. http://research.microsoft.com/pubs/155552/decisionForests_MSR_TR_2011_114.pdf
Last time I checked the literature (a few years ago now) there appeared to be two good first choices of image feature. SIFT or sparse Haar wavelets.
Have a look at Criminisi et al 2008 (http://research.microsoft.com/pubs/72423/Criminisi_bmvc2008.pdf) for a random forest and Haar wavelet based object recognition system.
An alternative approach from Fergus et al. 2007, (http://cs.nyu.edu/~fergus/papers/fergus_ijcv.pdf) uses a simple image patch model tied together using a Bayesian network.
OpenCV is probably as good place as any to start to find existing code. Matlab also claims to have good support for these tasks.