A learning-based human facial image quality evaluation method in video-based face recognition systems

Video-based face recognition systems suffer from severe performance degradation under uncontrolled real-world conditions. Using multiple images can enhance recognition performance, however, also introduces extra computational burden. In this paper, we propose a learning-based facial image quality evaluation method, which can be applied to selection of high-quality images and perform quality-based importance weighting in video-based face recognition systems. Features are carefully designed for human faces particularly, and are capable of applying in real-time applications for their low computational complexity. A random forest regressor is utilized to learn a subjective quality function, which is trained on a database labeled manually where the quality of images is scored from 1 to 5. Experiment demonstrates that the proposed method can effectively estimate the subjective quality score of facial images, and can lead to performance gain when applied in video-based face recognition systems.