Recognizability Embedding Enhancement for Very Low-Resolution Face Recognition and Quality Estimation

Figure 1. A deep face model pretrained on high-resolution face images introduces a cluster of unrecognizable instances (grey spike), dubbed unrecognizable identities (UIs) in [9]. (a) shows the bimodal distribution for a very low-resolution face dataset [6] based on distance against the UIs. Interestingly, a portion of hard-to-recognize faces (red peak) lie close to the UIs, indicating their low recognizability. (b)

We propose to improve the recognizability of hard-to-recognize instances by pushing them away from the UIs center. Consequently, faces with higher recognizability indexes are further apart from UIs center in the embedding space. Our method not only induces more discriminative representations but also translates face quality into a measurable indicator that closely matches human cognition.