Fingerprints and hand-written/printed characters processing methods for recognition via multi-stage self organized learning in a distributed computing environment.

摘要

In this thesis we present techniques for processing fingerprints, hand-written and printed characters for the purpose of their recognition using the Megherbi's coefficients in the M-coefficient space in a Distributed Computing Environment. Note that Megherbi's methodology and derivation of the M-coefficients are not the subject of this thesis and are beyond its scope. The main focus of this thesis is on the processing phase of objects before their recognition phase in the M-coefficients space. Particularly, we show how objects may undergo certain procedures before the M-coefficients are computed. In the case of fingerprint processing, we implement multiple steps processing before fingerprint feature extractions. While other methods use minutiae feature extraction, here based on Megherbi's methodology we represent the fingerprint entirely including minutiae information for accuracy purposes. We present techniques to extract fingerprints ridges, and cores. For the challenging extraction and segmentation of fingerprint cores, we introduce the concept of overlapped-window squared gradients directional fields. Through curve fitting models of the ridges we also show how to further improve fingerprints ridges processing accuracy. In the case of hand-written and printed character recognition, we show how proposed multidirectional morphological operators are applied to eliminate unwanted characters information and artifacts. We also present a technique based on the concept of image "Vertical Width" (VW) function for solving the challenging problem of hand-written character segmentation out of written words. Here the written words considered are those written by various individuals at different times. Character comparison is then performed in the M-coefficient space. We show the high hand-written character recognition accuracy we achieve using a two-stage self-organizing classifier, which results in similar hand-written/typed characters clustering into distinct groups in the M-coefficient space. Experimental results, advantages and computational performance analysis of the proposed processing methods are presented. Processing computations are implemented in the CMINDS HDPC (High Performance Distributed Computing) environment to demonstrate the ability to decrease processing time through the Message Passing Interface. Finally, the different steps of the work in this thesis are supported by the theories behind the proposed processing techniques before recognition in terms of proposed theorems, propositions, lemmas and their proofs of correctness.