Quality Assessment of Facial Images

Quality assessment of facial images differs from the traditional quality assessment of image and video signals with regards to its multiple goals such as to ensure its fidelity to the human visual system (HVS) model, to predict matching performance, to generate feedback on image acquisition, to guard the enrollment process, and to provide a weight for merging multimodal biometrics. In this paper, we present a quality assessment framework that complies with the requirements of ISO/IEC 19794-5 for facial biometrics and additionally ensures optimal recognition performance. This framework employs a novel classification-based score normalization process for various quality metrics and includes techniques to fuse those individual quality scores into an overall quality score which is shown to be correlated to the genuine match scores of the Facelt face recognition engine. We confirm the effectiveness of this overall quality score at satisfying multiple goals by first parameterizing ROC curves with average database quality to show the predictive nature of the metric and secondly by showing the consistency between this overall quality score and human perception of image quality.