In CT or MR images of the head, we extract typically 3000 to 5000 extremal points, which is much more than the 100 to 200 frames modeling the amino acids in a protein. Moreover, some crest lines are well conserved and provide a structure that we can exploit to reduce the complexity of the matching process.
We consider thus each crest line of the model image as an independent structure and index the couples of extremal frames on this line in the hash table. At recognition time, we try to match each crest line of the scene image with all the crest lines of the model using the algorithm previously described. Thus, we end up with matches of crest lines between the model and the scene, each match having an associated rigid transformation. To find global matches, we just have to run once again our clustering algorithm on transformations to find out the compatible line matches.