Face Super-Resolution Quality Assessment Based on Identity and Recognizability

Face Super-Resolution (FSR) plays a crucial role in enhancing low-resolution face images, which is essential for various face-related tasks. However, FSR may alter individuals' identities or introduce artifacts that affect recognizability. This problem has not been well assessed by existing Image Quality Assessment (IQA) methods. In this paper, we present both subjective and objective evaluations for FSR-IQA, resulting in a benchmark dataset and a reduced reference quality metrics, respectively. First, we incorporate a novel criterion of identity preservation and recognizability to develop our Face Super-resolution Quality Dataset (FSQD). Second, we analyze the correlation between identity preservation and recognizability, and investigate effective feature extractions for both of them. Third, we propose a training-free IQA framework called Face Identity and Recognizability Evaluation of Super-resolution (FIRES). Experimental results using FSQD demonstrate that FIRES achieves competitive performance.