UG-LDFace: Unified and Generalized Framework for Long-Range Disguised Face Recognition

Muskan Dosi, Chiranjeev Chiranjeev, Richa Singh, Mayank Vatsa

Abstract: Long-range, low-resolution videos have widespread applications in active monitoring, crowd counting, traffic analysis, and person verification/identification. The problem of analyzing faces in such an environment is exacerbated by the presence of disguise and occlusion. This research presents UG-LDFace, a novel face recognition model to address this arduous challenge. A single-stage unified framework is proposed that integrates two feature refinement techniques: feature enhancement for low-resolution data and feature selection for disguised faces. The proposed model also comprises a revised distribution technique to generalize UG-LDFace on unseen data. The proposed approach shows its efficacy on five different datasets, DroneSURF, SCface, D-LORD, DSIMF, and LFW, containing various levels of occlusion and low-resolution data.