Section: Partnerships and Cooperations

National Initiatives

ANR Attelage de systèmes hétérogènes

Participants : Guillaume Gravier, Bogdan Ludusan.

Duration: 3 years, started in November 2009.

Partners: IRISA, LIA, LIUM

The project ASH (Automatic System Harnessing – ANR-09-BLAN-0161-03) aims at developing new collaborative paradigms for speech recognition. Many current ASR systems rely on an a posteriori combination of the output of several systems (e.g., confusion network combination). In the ASH project, we investigate new approaches in which three ASR systems work in parallel, exchanging information at every step of the recognition process rather than limiting ourselves to an a posteriori combination. What information is to be shared and how to share such information and make use of it are the key questions that the project is addressing. The collaborative paradigm is being extended to landmark-based speech recognition where detection of landmarks and speech transcription can be considered as two (or more) collaborative processes.

ANR FIRE-ID

Participants : Sébastien Campion, Philippe-Henri Gosselin, Patrick Gros, Hervé Jégou.

Duration: 3 years, started in May 2012.

Partner: Xerox Research Center Europe

The FIRE-ID project considers the semantic annotation of visual content, such as photos or videos shared on social networks, or images captured by video surveillance devices or scanned documents. More specifically, the project considers the fine-grained recognition problem, where the number of classes is large and where classes are visually similar, for instance animals, products, vehicles or document forms. We also assumed that the amount of annotated data available per class for the learning stage is limited.

ANR Secular

Participants : Laurent Amsaleg, Teddy Furon, Benjamin Mathon, Ewa Kijak.

Duration: 3 years, started in September 2012.

Partners: Morpho, Univ. Caen GREYC, Telecom ParisTech, Inria Rennes

Since their invention, content based image retrieval systems (CBRS) and biometric systems have evolved separately. This is due to the fact that they originate from different research and industrial communities. This Basic Research project, called SecuLar, groups researchers from both communities who have observed that both type of systems have indeed a lot in common in terms of goals and technological blocks. These techniques are used, however, in quite different settings possibly explaining the gap between the two. The people involved in this SecuLar project believe that what is specific to each family of approach can now benefit the other for the two following fundamental reasons.

Biometrics needs scale. The size of biometric databases quickly increases. It grows in terms of the number of records kept in the database. It also grows in terms of the size of each record as larger biometric templates maintain high quality recognition. The amount of data becomes large enough to require powerful indexing techniques. CBRS are good at this as they allow ultra fast searches of nearest neighbours in huge datasets. But porting these techniques to a biometric context is far from being easy. Biometric databases are typically protected to enforce confidentiality and privacy as security is paramount. Indexing biometric data is thus difficult because the techniques enforcing security in biometrics conflict with the technique bringing efficiency to database searches. No biometric system can today cope with both all the privacy and security constraints and the scale at which they should work in the real world for new applications.

CBRS need security and privacy. We witness a new use of CBRS these days. CBRS become the main multimedia security technology to enforce copyright laws (content monetization) or to spot illegal contents (detection of copies, paedophile images, ...) over the Internet. However, they were not designed with privacy, confidentiality and security in mind. This comes in serious conflict with their use in these new security-oriented applications. Privacy is endangered due to information leaks when correlating users, queries and the contents stored-in- the-clear in the database. It is especially the case of images containing faces which are so popular in social networks. Biometrics systems have long relied on protection techniques and anonymization processes that have never been used in the context of CBRS. Here, we plan to understand how biometrics related techniques can help increasing the security levels of CBIRS while not degrading their performance.