Optimizing Super Resolution for Face Recognition

Face Super-Resolution is a subset of Super Resolution (SR) that aims to retrieve a high-resolution (HR) image of a face from a lower resolution input. Recently, Deep Learning (DL) methods have improved drastically the quality of SR generated images. However, these qualitative improvements are not always followed by quantitative improvements in the traditional metrics of the area, namely PSNR (Peak Signal-to-Noise Ratio) and SSIM (Structural Similarity Index). In some cases, models that perform better in opinion scores and qualitative evaluation have worse performance in these metrics, indicating they are not sufficiently informative. To address this issue we propose a task-based evaluation procedure based on the comparative performance of face recognition algorithms on HR and SR images to evaluate how well the models retrieve high-frequency and identity defining information. Furthermore, as our face recognition model is differentiable, this leads to a novel loss function that can be optimized to improve performance in these tasks. We successfully apply our evaluation method to validate this training method, yielding promising results.