Tesis doctoral de Enric Sesa Nogueras
The main motivation of this dissertation is the exploration of the field of online text-dependent writer recognition, in order to provide evidence of the usefulness of short sequences of text to perform identification and verification, which are the two tasks involved in recognition. From this motivation stem its main goals and contributions: an exploration performed from a practical perspective, thus requiring the development of a recognition system, and the gathering of evidence concerning the discriminative power of in-air trajectories (the trajectories described while not exerting any pressure on the writing surface, when the hand moves in the air while transitioning from one stroke to the next), i.E. Their ability to discriminate among writers. In-air and on-surface trajectories have been analyzed from the perspective of information theory and the results yielded by this analysis show that, except for pressure, they contain virtually equal amounts of information and are notably non-redundant. This suggests that in-air trajectories may have a considerable discriminative power and that they may help improve the overall recognition performance when combined with on-surface trajectories. An innovative writer recognition system that fulfils the abovementioned practical goal has been devised. It follows an allographic approach, that is, it does not take into account the global characteristics of the text but focuses on character and character-fragment shapes. Strokes are considered the structural units of handwriting and any piece of text is regarded as two separate sequences, one of pen-up and one of pen-down strokes. The system relies on a pair of catalogues of strokes, built in an unsupervised manner by means of self-organizing maps, which allow mapping sequences of strokes into sequences of integers. The latter sequences, much simpler than the original ones, can be effectively compared by means of dynamic time warping, which takes advantage of the neighbouring properties exhibited by self-organizing maps. Measures obtained from each sequence can be combined in a later step. The recognition system has been experimentally tested using 16 uppercase words from the biosecurid database, which contains 4 executions of each word donated by 400 writers. The experimental results obtained clearly sustain the claim that online words have a notable recognition potential and show the suitability of the allographic approach to perform writer recognition in the online text-dependent context. Regarding identification, the system compares positively to other word-based identification schemas. As for verification, the accuracy levels attained do not lie much below the accuracies reported for today¿s state-of-the-art signature verification methods. Furthermore, the results obtained from in-air trajectories have substantiated what the information analysis had already suggested: their considerable recognition power and their notable non-redundancy with respect to on-surface trajectories. Finally, a new method to generate synthetic samples of online words from real ones has been proposed. This method is based on the recognition system previously described, takes advantage of its main characteristics and can be seamlessly integrated into it. Synthetic samples are used to enlarge the enrolment sets, which has the effect of substantially improving the recognition accuracy of the system.
Datos académicos de la tesis doctoral «An online writer recognition system based on in-air and on-surface trajectories«
- Título de la tesis: An online writer recognition system based on in-air and on-surface trajectories
- Autor: Enric Sesa Nogueras
- Universidad: Politécnica de catalunya
- Fecha de lectura de la tesis: 20/09/2012
Dirección y tribunal
- Director de la tesis
- Marcos Faundez Zanuy
- Tribunal
- Presidente del tribunal: pedro Gómez vilda
- jordi Sole casals (vocal)
- (vocal)
- (vocal)