Physionomie

The research interest for people faces identification, as part of judicial investigations, is well established and remains one of the key elements of success. However, the technologies used in uncontrolled conditions of shooting are still in their early stages of development and have yet to be improved and validated in real live test in order to make them fully operational. Given some recent progress, the recognition task is difficult because face images extracted from video-protection systems are poorly resolved, poorly lit and variably oriented. The PHYSIONOMIE project proposes to develop new tools of physiognomic recognition by: • improving the images of inquiry (pictures or video clips), • exploiting 3D face models to change the orientation of sought faces and then obtaining images with comparable exposures, • defining new similarity measures on partially comparable data (query and reference images are of different nature, presence of expression, occlusion, ...). The accuracy of biometric recognition algorithms depends on the quality of image acquisition and its overall performance depends greatly on the database size used. The creation of a 3D face images database of significant size would require a substantial time. Thus, PHYSIONOMIE will develop tools to build 3D face image models from anthropometric pictures. The validation of this approach will be done through 3D face models produced from low cost sensors such as stereoscopic smartphones, Kinect sensors, etc.., on face pictures acquired during the project. Physiognomic recognition tools will provide as a result a face list ranked in order of relevance, with the sought face that does not always appear at the first rank. Therefore, in PHYSIONOMIE, we propose to incorporate within the demonstrator intuitive and ergonomic capabilities of visual search to navigate through the results list. We will use the new similarity measures as well as soft biometrics (eye colour, presence of scars / tattoos ...) to navigate through the global database from one or more faces of the list returned. End-users, partners of the project, are willing to test and evaluate these tools. In this context, the project will measure the tool potential and will aim to evolve the similarity measure technologies for forensic identification. It will adapt the technology to operational needs, will specify the conditions of use and will measure the gain in efficiency. The possible deployment of a tool with PHYSIONOMIE features in operational services cannot be done without studying the societal acceptability of this technology and the project will conduct a qualitative and quantitative study in this area. Its findings will complement the technology and deployment roadmap, and will specify the contexts of use that are acceptable to society.