Image matching is a fundamental problem in computer vision. In this thesis, we address the image matching problem as learning and classifying correspondences. More precisely, we formulate the image matching problem as: given a set of training pairs of images that implicitly captures the transformation(with both positive and negative classes), identify if a new pair of test images is matched via the transformation class. In this formulation, all the training data, as well as test data, are image pairs. The approach taken is to consider only relative visual content, rather than absolute visual content, so the learned image matching classifier could be applied to images of totally different visual content as compared to the training data. This is in contrast to appearance-based object detection methods, for which once the training process is completed, the classifiers may only be used to recognize objects of the same categories with the training images.