Yüz ve imza bilgileri kullanarak çok kipli kullanıcı doğrulama

Biometric verification systems are widely used to verify the identity of a person. Face, signature, fingerprint and iris are among popular biometrics. They are used in a wide range of applications such as security, shopping and finance. It is required them to have very low error rates especially in forensic applications. They have to deal with several problems to obtain acceptable results. In order to overcome limitations of unimodal biometric systems, a multimodal verification system is presented. Signature and face traits are used to build unimodal biometric systems. Then, score level combination of these system is utilized to reach lower error rates. Several attack and noise procedures are applied to evaluate performance of the multimodal verification system. The usage of recurrent and convolutional neural network architectures, that have achieved great success in a broad range of computer vision tasks, are investigated for offline signature verification. It is shown that, combinations of these two different approaches can be used to achieve state of the art results. User-independent and user-dependent approaches are investigated to perform authentication. Two convolutional neural network architectures are deployed to learn user-independent signature features. Then, user-dependent classifiers are trained to accept or reject an identity claim. A transfer learning approach is utilized to develop a face verification system. User-independent face features are extracted from a pre-trained convolutional neural network. Then, these features are fed into user-dependent classifiers to perform verification. In this thesis, face and signature verification systems are developed and their performances are evaluated separately. Then, a multimodal biometric system, which fuses information coming from two biometrics, is proposed. Results show that, multimodal approach can be used to obtain higher accuracy than unimodal systems and make the system robust against spoof attacks and noise.