FIQ-Quest - Exploration and Exploitation of Probabilistic Interpretable Face Image Quality

• Applicant: Dr.-Ing. Philipp Terhörst
• Subject Area: Image and Language Processing, Computer Graphics and Visualisation, Human Computer Interaction, Ubiquitous and Wearable Computing ; Security and Dependability, Operating-, Communication- and Distributed Systems
• Term: since 2025
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 544631027

Face Image Quality (FIQ) is defined as the utility of a face image for the purpose of recognition prior to any matching. Images of low FIQ are likely to produce wrong matching decisions. Depending on the application of the face recognition systems, this can come at high financial or societal costs. Thus, it is of high importance that face images given to a face recognition system are of sufficient quality. Previous works on FIQA Assessment (FIQA) are built on a fuzzy definition of FIQ resulting in (a) low generalizability, (b) low interpretability, and (c) low reusability. The low generalizability refers to unstable performances across datasets and face recognition models, which limits their success in real-world applications. The low interpretability refers to the fact that no clear relation of the quality values to specific matching error rates is given, which does not provide any matching guarantees and makes it hard for system operators to work with. The low reusability refers to the difficulty of integrating FIQ into other tasks of a biometric system. Due to the lack of interpretability and the fuzzy definition of FIQ, it is challenging to incorporate this valuable source of information into related tasks, such as detecting attacks on biometric systems. This project will fix these shortcomings in three steps through exploration and exploitation. First, it aims to explore new definitions of FIQ that connect the quality of two samples with the probability that the matching decision of the two samples is correct. Unlike the current FIQ definition, the new definition will allow interpreting quality estimates of how well a sample will perform for recognition. The second step will build FIQA solutions and evaluation metrics based on these new definitions to achieve more generalizable and interpretable quality estimation. In the third step, the developed solution will be used to effectively recognize identities and attacks on face recognition systems. This will exploit the developments of the previous exploration steps to ensure a wider impact of the developed solutions. To secure social sustainability and acceptability of the developed solutions, the project further investigates different strategies to include FIQ in a face recognition system to ensure that all groups of the population are treated fairly by the system and that they are perceived as such. At the end of the project, guidelines for future works in this direction will be developed and published based on the lessons learned to support a socially and economically sustainable development of face recognition technology jointly benefiting industry, governance, and society.

• DFG Programme: Research Grants
• International Connection: Norway, Slovenia
• Cooperation Partners: Professor Dr. Kiran Raja; Professor Dr. Raghavendra Ramachandra; Professor Dr. Vitomir Struc