Alumno: Juan José Igarza Ugaldea
Directores: Inma Hernáez e Iñaki Goirizelaia
Fecha defensa: 12 mayo 2006
Day by day it is becoming more important to find a simple, easy-to-use, fast and convenient way to verify people’s identity. One way to do this is the long-traditional, socially approved biometric identification procedure known as the signature. This dissertation contains two proposals for the biometric recognition of written signatures.
For the purposes of this study, a substantial number of signers and a number of quality forgeries were collected in a multimodal biometric data-base, which included fingerprints and voice recordings as well as signatures by anonymous signers and skilled forgeries.
A study of the current state of the field of biometry was also carried out, first reviewing the field in general terms, then moving on to look at handwriting analysis, and finally examining the two aforementioned forms of written signature.
The proposal for on-line signature recognition is a system referring to global parameters (centre of the ink mass and main lines of inertia) and based on local features (coordinates, speed, accelerations, pressure and pen angle). I propose a simple scaling algorithm that is valid for all the local features to standardise signature time. For a point of equilibrium between level of security and processing and storage capacities, I limited processing with Hidden Markov Models (HMM) to nine local features, obtaining results comparable to co-error rates found in a research situation with six-state LR-HMMs.
I studied a reference system to improve signature alignment. After proving that using the coordinates of the starting point of dynamic signatures and slant is a source of noise, I proposed a reference system located at the centre of the ink mass and oriented to ink inertia lines. The new reference system achieves better signature alignment, while using only the coordinates when known leads to substantial improvement in the co-error rate.
For off-line signature recognition, there are two systems fed by spatially ordered sequences of points and “”pseudo-dynamic”” geometrical features derived from those sequences of points, using the dynamic technique of LR-HMMs appropriate for analysis of dynamic features. The purpose of these proposals is to extend LR-HMMs to off-line signature recognition through connectivity analysis. Connectivity analysis classifies image components into similar adjacent sets of pixel called blobs.
Two criteria for creating models according to the way of organising the blobs given by connectivity analysis are compared. The order criterion for the first method proposed is perimeter length; in the second method, the blobs are arranged in their natural order. Results obtained in a number of experiments are presented, showing that the proposed methods obtain similar rates to those found in the research situation. The models based on reading order are better adapted to the Latino signature culture.