Low-resolution face recognition and the importance of proper alignment

Face recognition methods for low resolution are often developed and tested on down-sampled images instead of on real low-resolution images. Although there is a growing awareness that down-sampled and real low-resolution images are different, few efforts have been made to analyse the differences in recognition performance. Here, the authors explore the differences and demonstrate that alignment is a major cause, especially in the absence of pose and illumination variations. The authors found that the recognition performances on down-sampled images are flattered mostly due to the fact that the images are perfectly aligned before down-sampling using high-resolution landmarks, while the real low-resolution images have much poorer alignment. To obtain better alignment for real low-resolution images, the authors apply matching score-based registration which does not rely on accurate landmarks. The authors propose to divide low resolution into three ranges to harmonise the terminology: upper low resolution (ULR), moderately low resolution (MLR), and very low resolution (VLR). Most face recognition methods perform well on ULR. MLR is a challenge for commercial systems, but a low-resolution deep-learning method can handle it very well. The performance of most methods degrades significantly for VLR, except for simple holistic methods which perform the best.