| Type de document | Thèse |
| Langue | eng |
| Titre | Off-line handwritten signature verification using the contourlet transform [ressource textuelle, sauf manuscrits] |
| Auteur(s) | Hamadene, Assia (Auteur) Chibani, B. (Directeur de thèse) Université des sciences et de la technologie Houari Boumediène (Editeur (scientifique)) |
| Adresse bib. | Alger : USTHB,2017 |
| Collation | 122 p. : ill. ; 30 cm + CD-Rom |
| Notes | Bibliogr. p. 114-122 |
| Notes de thèse | Doctorat : Réseaux et télécommunications : Faculté d'Electronique et Informatique : Université des sciences et de la technologie Houari Boumediène : 2017 |
| Theme | Electronique |
| Mot (s) clé | Signatures électroniques Systèmes de télécommunications |
| Résumé | Among biometric systems, the handwritten signature verification is one of the most complex biometric applications because the verification is based on the analysis of the handwritten behavioral action. The main dilemma is that, on one hand, the behavioral aspect of handwriting is characteristically specific to each writer and, on the other hand, the relevancy of automated system lies on its generalized applicability to all writers. Moreover, a high similarity between two signatures does not necessarily mean that they have been written by the same person. In fact, this case can occur when the signature had been skillfully reproduced by another person. Conversely, a low similarity between two signatures does not necessarily mean that it comes from two different writers because of the intra-writer variability. The signature analysis can, therefore, turn into an extremely complex problem requiring different disciplines to be involved. Hence, this research is focused on developing new feature extraction methods using Contourlet transform to build robust handwritten signature verification systems. The main issue is to extract the most discriminative property which allows to efficiently discriminate between genuine signatures and highly skilled forgeries. In this context, we proposed mainly four contributions to deal with the most challenging tasks of HSVs which can be summarized as follows: • Two feature generation methods are developed, namely, Normalized Energies and Directional Code Co-occurrence Matrix (DCCM), both of them is based on directional information contained into the signature deduced from the CT. Normalized Energies method exploits the level of information according to each direction contained into the signature. Whereas, the DCCM features exploit the quantitative and spatial directional information to characterize the handwritten signature, where The main idea is to attribute to each signature segment the corresponding code of its dominant direction. Then, the resulting directional coded image is analyzed in terms of localization and occurrences of signature directions using co-occurrence matrix. • A whole HSV system using DCCM features is developed to address the Writer-Independent (WI) HSV concept using a new one-class dissimilarity based protocol, In this context, we proposed a new stability criterion as leaning stage to define a Writer independent threshold rather than using classical machine learning. A new concept of WI learning is further proposed through mixing databases which improves the effectiveness of the proposed learning criterion as well as classification performances. The protocol uses a reduced number of genuine reference signatures conversely to usual writer-independent approaches which use a binary classifier as well as a large number of references and allows to achieve verification independently of a selected dataset model. • Finally, in order to improve the DCCM performances, a new feature generation method, we call, Local Directional coded pattern (LDCP) is developed which is mainly based on DCCM feature generation steps and aims to construct directional structure from the DCCM directional map taking into consideration a larger neighborhood and generates a unique code for each directional structure contained into the signatures to analyze the signature directional distribution. |
Hamadene, Assia
Off-line handwritten signature verification using the contourlet transform [ressource textuelle, sauf manuscrits] / Assia Hamadene; Dir. B. Chibani; Ed. Université des sciences et de la technologie Houari Boumediène.-Alger : USTHB,2017.-122 p. : ill. ; 30 cm + CD-Rom.
- Bibliogr. p. 114-122
Doctorat : Réseaux et télécommunications : Faculté d'Electronique et Informatique : 2017
.
Signatures électroniques
Systèmes de télécommunications
Among biometric systems, the handwritten signature verification is one of the most complex biometric applications because the verification is based on the analysis of the handwritten behavioral action. The main dilemma is that, on one hand, the behavioral aspect of handwriting is characteristically specific to each writer and, on the other hand, the relevancy of automated system lies on its generalized applicability to all writers. Moreover, a high similarity between two signatures does not necessarily mean that they have been written by the same person. In fact, this case can occur when the signature had been skillfully reproduced by another person. Conversely, a low similarity between two signatures does not necessarily mean that it comes from two different writers because of the intra-writer variability. The signature analysis can, therefore, turn into an extremely complex problem requiring different disciplines to be involved. Hence, this research is focused on developing new feature extraction methods using Contourlet transform to build robust handwritten signature verification systems. The main issue is to extract the most discriminative property which allows to efficiently discriminate between genuine signatures and highly skilled forgeries. In this context, we proposed mainly four contributions to deal with the most challenging tasks of HSVs which can be summarized as follows:
• Two feature generation methods are developed, namely, Normalized Energies and Directional Code Co-occurrence Matrix (DCCM), both of them is based on directional information contained into the signature deduced from the CT. Normalized Energies method exploits the level of information according to each direction contained into the signature. Whereas, the DCCM features exploit the quantitative and spatial directional information to characterize the handwritten signature, where The main idea is to attribute to each signature segment the corresponding code of its dominant direction. Then, the resulting directional coded image is analyzed in terms of localization and occurrences of signature directions using co-occurrence matrix.
• A whole HSV system using DCCM features is developed to address the Writer-Independent (WI) HSV concept using a new one-class dissimilarity based protocol, In this context, we proposed a new stability criterion as leaning stage to define a Writer independent threshold rather than using classical machine learning. A new concept of WI learning is further proposed through mixing databases which improves the effectiveness of the proposed learning criterion as well as classification performances. The protocol uses a reduced number of genuine reference signatures conversely to usual writer-independent approaches which use a binary classifier as well as a large number of references and allows to achieve verification independently of a selected dataset model.
• Finally, in order to improve the DCCM performances, a new feature generation method, we call, Local Directional coded pattern (LDCP) is developed which is mainly based on DCCM feature generation steps and aims to construct directional structure from the DCCM directional map taking into consideration a larger neighborhood and generates a unique code for each directional structure contained into the signatures to analyze the signature directional distribution.