Hallucinating faces in the dark
Sasikanth Bendapudi, Khoa Luu, Marios Savvides · 2012
Correction of uneven illumination in face images has been a task of intense research with its applicability being extensive. Most algorithms look to improve recognition accuracy failing to produce visually good results or viceversa. In this work, we present a data-driven illumination correction algorithm which simultaneously produces visually good results and improves recognition accuracy among faces. The illumination behavior on the generic structure of faces is learnt by training pyramid sub-spaces on illumination images. An algorithm is proposed to iteratively project illumination information onto these sub-spaces and filter out other information projected onto them. Results are presented on the CMU-PIE and Extended Yale-B databases. Performance is measured using PSNR measures for relighting images and improvement in the performance of face recognition rate in comparison to other contemporary performers.