Face Detection and Recognition Based on General Purpose DNN Object Detector

Facial recognition is a very on-demand task today. What makes a racial detection and recognition system good is its robustness to changes in illumination, rotation and tilt, position and occlusion. In this paper we present a face recognition system trained on a proprietary datatabase. Our method is based on YOLO (You Only Look Once) model, which is a general object detection system, that by modifying and testing different internal parameters we have proven it to be adequate for face recognition. The database was created especially for this project, with images featuring 10 different subjects, different orientations, and illumination conditions, totalling to 120.000 samples. The models were based on 17 variations of the architecture of Darknet-19 neural network, from which we have chosen the one with the best results.