See this page for a comparison of classification methods : http://www.mathworks.com/machine-learning/index.html
For training set, I'd pick images that span the entire range of what you ever expect to encounter, from typical case to real extreme cases (whatever that might be). If you don't train on data near the edges of your range, then the classifier might not be very good out there. You don't want to train on just images near the middle or typical/average images or else the classfication mgiht not be very good when you encounter an extreme case. There is nothing wrong with hand picking images to make sure you have a complete set of representative images that totally span the range.
