Partners
University of Bern, Institute of Informatics and Applied Mathematics (IP leader):
Horst Bunke, http://www.iam.unibe.ch/~fki/
ETHZ/TIK, Speech Processing Group: Beat Pfister, http://www.tik.ee.ethz.ch/
IDIAP, Herve Bourlard, Samy Bengio, http://www.idiap.ch/

Context and Goals
An important issue in multimedia access and content protection is the verification of enrolled users. The focus of this individual project is on the automatic identification and authentication of persons through the use of biometric features. In its first phase the project will put emphasis on methods that use a single modality only, for example, fingerprints, voice, or facial images. Later the integration of different biometric features into single multimodal identification systems will be studied.

Automatic person identification and verification is interesting in the context of IM2 for two reasons. First, IM2 aims at the development of complex multimodal information systems. Such systems often incorporate critical information, access to which has to be restricted to certain individuals. Therefore, access control mechanisms and in particular the automatic identification of persons is a crucial part of any advanced multimodal information system. Secondly, many different biometric features for identity verification have been proposed. They include, among many others, fingerprints, facial images, voice and signature. There is a consensus that through the use of different sensor modalities and decision procedures the robustness and reliability of a system can be improved. The integration of multiple information modalities is one of the key issues in IM2. Hence development of identity recognition and verification systems based on multiple biometric features is a topic central to IM2.

There are close links to a number of other IM2 projects, in particular to IP.SA where facial image analysis is studied, and to IP.SP where the focus is on speech analysis.

Research Issue
Identification of individuals based on biometric features has become a very active area of research recently. In the first phase of this project research will focus on biometric features and person identification based on the following single input modalities:

• Identification and recognition of faces in images (IDIAP): The main objective is to investigate effective methods for the identification or verification of people based on facial characteristics.
• Fingerprints (Uni Bern): The problem considered here is the verification, classification and recognition of fingerprints for the identification of persons.
• Speaker verification and recognition (ETHZ and IDIAP): The problem to be solved is the identification of users of a multimodal information system based on his or her voice.

At a later stage of this project the integration of various biometric features will be studied. Possible solutions to this problem include the fusion of features from various modalities and their use in a single classifier or the application of multiple classifier architectures, where information fusion takes place at the decision level (i.e. at the classifiers' output level).

Outcomes
Year 1:

• Fingerprints (Uni Bern)

- Survey existing methods for fingerprint verification, classification and recognition, including image preprocessing, segmentation, and feature extraction. Compile a list of publicly available fingerprint databases and publicly available software tools for fingerprint analysis.

- Develop methods for fingerprint image preprocessing, segmentation, and feature extraction. Acquire a database of fingerprint images.

• Speaker verification and recognition (ETHZ and IDIAP)

- A speech signal database that allows to evaluate the decision reliability of text-dependent speaker verification algorithms, particularly as a function of utterance duration, noise level, speaker's sex, telephone type, etc. This database will enable all subsequent investigations.

- Adaptation of an existing speaker verification algorithm to the TORCH platform to allow easy and flexible integration with other modalities and fusion algorithms.

- Investigation of the decision reliability as a function of utterance duration, signal quality, telephone type, etc.

• Face recognition and identification (IDIAP)

- Software development for robust face detection and verification.

• Modality fusion (IDIAP)

- Implementation and evaluation of several fusion algorithms.

• Fingerprints (Uni Bern)

- Build prototypes of fingerprint verification, classification, and verification systems, based on generic software modules for graph matching.

• Speaker verification and recognition (ETHZ and IDIAP)

- Development of a new method (based on a neural network) to measure the difference between two speech signals with the aim to achieve a maximum discrimination between the situations of same speaker and different speakers.

- A comparison of the error rates of different approaches to speaker verification.

• Face recognition and identification, and modality fusion (IDIAP)

- Optimization of face detection and verification systems; testing on evaluation database.

- Development and adaptation of a multi-stream approach, in the framework of TORCH, towards multi-modal user authentication.

Quarterly status reports
Available on the local site (password protected).