Demographic Variability of Face Image Quality Measures in Operational Data

Face image quality assessment plays a critical role in operational face recognition systems, where quality thresholds directly affect recognition performance, throughput, and fairness. This work presents a large-scale empirical analysis of demographic variability in the ISO/IEC 29794-5-compliant Open Source Face Image Quality (OFIQ) measures using real operational data from automated border control scenarios. The dataset comprises more than a hundred of thousand constrained reference images and corresponding live capture images acquired at border gates, enabling a realistic assessment of OFIQ behavior under deployment conditions. The unified quality score and multiple component quality measures are analysed across gender, age groups, skin tones and region-of-origin. The results reveal pronounced demographic variability in the unified quality score of OFIQ as well as several component quality measures, in particular the ones related to luminance, sharpness and contrast. It is further shown that this variability leads to substantial differences in percentile-based quality thresholds across demographic subgroups. Definitions of globally applicable thresholds is challenging for operators but suggestions are provided. In addition, we identify saturation effects caused by normative scalar mappings, limiting their discriminative power in operational settings. The study provides actionable insights for system operators and informs ongoing standardization and algorithmic efforts, including the development of OFIQ 2.0.