Methodological insights on restoring face photos of multinational passports

We study the problem of restoring severely degraded face images scanned from multinational passport photos. The purpose of this paper is to illustrate alternative viable image restoration strategies aiming to improve face recognition performance in such a realistic scenario. The contributions of this work are two-fold. First, the West Virginia University (WVU) Passport database, composed of face images of various subjects, is collected at WVU as a part of a CITeR project¹. These images illustrate the challenges associated with matching severely degraded face images. Second, various pre-processing schemes are tested and a final strategy is proposed in order to eliminate the noise present in degraded face images. After conducting an extensive experimental study, it is established that the proposed restoration scheme shows promise in improving the overall image quality of passport photos as well as heterogeneous face recognition performance (i.e. rank-1 identification accuracy). This conclusion is drawn after analyzing the results from matching the restored passport images (probe dataset) against the ones that are of good quality (gallery dataset).