You are going to need to create feature vectors from the images and distinguish based upon those.
SVM is traditionally a two-class process (multi-class takes a lot more memory and time), and finds a dividing line between two sets of data. In some cases that dividing line could be an arc instead of a straight line, but it still comes down to taking the data and saying "this is on one side and the other side is the other class". You are certainly not going to be able to do that if your data being classified is the images themselves.