Thesis (M.Sc.) - Cairo University - Faculty of Computers and Artificial Intelligence - Department of Computer Science

One of the most important fields of computer vision has been Face Recognition (FR). Face recognition applications are utilized to recognize images of faces from videos captured across several shared security cameras. The issue of facial recognition can be classified into 2 categories, the first one is recognition with more than one reference per individual, which can be referred to as a traditional facial recognition issue. The other one is recogni- tion using only one reference per individual. The performance of facial recognition models decreases due to insufficient references in particular Single Sample Per Person (SSPP) and faces that captured in the Operational Domain (OD) different from faces captured in the Enrollment Domain (ED) in lighting, low resolution and pose. This thesis proposed a sys- tem that would address all issues related to FR with SSPP. 3D face rebuilding is utilized to augment the reference set with various poses and to create a design domain dictionary to beat the limited reference issue. In addition, the design domain dictionary is utilized to feed various deep neural networks. Face lighting transfer methods are used to beat the issue of lighting. Labeled Faces in the Wild database (LFW) is utilized for training the Super-Resolution Generative Adversarial Network (SRGAN) to beat the issue of a low resolution. The LFW database is used for training Deblur Generative Adversarial Net- work (DeblurGAN) to beat the issue of blurriness.The proposed system evaluated using the Chokepoint database and the COX-S2V database. The proposed system with transfer learning (FaceNet) gives the highest accuracy up to 98.5% on the COX-S2V database and 98.7% on the Chokepoint database. The final results confirm an overall increase in accu- racy especially in comparison to the methods used by SSPP for face recognition (generic learning method and face synthesizing method). The proposed system also exceeds the accuracy of the Traditional and Deep Learning (TDL) method, which uses SSPP for facial recognition

Issued also as CD