Deep multi-frame face hallucination for face identification.

Face verification problem has been successfully approached in many recent papers. Unfortunately, recognition quality decreases significantly in the presence of different degradation factors inherent to real world data, for example surveillance videos, where faces can be very small in size, blurred and affected with compression artifacts. Here we present face hallucination approach based on using several sequential video frames: our neural network restores the central frame of each input sequence additionally taking into account a number of adjacent frames, warped to align with the central one. We show that our face hallucination approach can improve face identification quality compared to the single-frame baseline and to the multi-frame approach without adjacent frame warping.