Similitude invariant pattern recognition on technical documents

The problem of geometric invariant pattern recognition is addressed. Specifically, we investigate the use of the Fourier-Mellin Transform (FMT) for the recognition of characters and symbols on technical documents, issued from technical paper maps of the French telephone network. In these documents, characters and symbols are drawn at any orientation and scale factor. Besides this specificity, a difficulty arises when characters and symbols remain connected after the binarization step. We show here that the FMT-based recognition tool that we propose performs well, either in terms of correct classification rates, when compared with the Zernike moments on large data sets, and also in solving the problem of detection and recognition of overlapping characters