- Sven Loncaric, project leader
- Marko Subasic
- Adam Hedji
- Josef Birchbauer, Siemens, Austria
- Faculty of Electrical Engineering and Computing, University of Zagreb, Croatia
- Siemens AG, Austria
The project was funded by Siemens, Austria.
The project duration was from 2005-2007.
Modern biometric passports contain different biometric data, but they all contain ID photographs. Such photographs are intended for automatic face recognition and validation which require image quality sufficient to support recognition methods. The quality level is estimated by checking a number of predefined requirements that ID images have to satisfy (defined by ICAO standard). Besides manually by human operator, such testing of images can be done automatically to some extent. Simpler requirements can easily be done automatically, but the more complex tests require knowledge of the scene composition. Once appropriate regions of interest in ID photo are identified it is easy to check more complex quality requirements. Examples of complex requirements are: no hair over eyes, face dimensions compared to image dimensions, no red eyes etc. The difficult part is to to find objects of interest in ID photographs.
The goal of the project was to segment input ID image into five most common regions. The most significant problem came from the fact that segmentation was a perquisite for quality check and hence hat to perform well on poor quality images (poor quality would be detected subsequently). The difficult problem required a substantial use of prior knowledge of the problem. Several knowledge based approaches have been tried: expert systems, neural networks, boosted classifiers, graph cuts etc. for training of learning based approaches, a image database has been formed with careful selection of images to include all possible variations in ID images. The final result of the project was a prototype software application.
The complete list of publications is available here.