Real Time Multiple Face Recognition: A Deep Learning Approach

Abstract

Though a lot of research has already been done in the field of Face Recognition, one amongst the remaining challenges is recognizing multiple faces in weird conditions in a large group size. A robust face recognition system has been developed which detects faces in multiple, occluded, posed images obtained under low illumination conditions. The detector is a trained 34 layered Residual Network which obtains an accuracy of 98.4% on Visual Geometry Group Dataset [1]. A hybrid model has been proposed by combining the Residual Network detector with the novel approach of face embedding using triplet loss function [2] for recognition. The numerical and graphical results attached in the report depict the effectiveness of the proposed model for a variety of conditions. A 22 layered Inception Network has been trained for feature extraction and it achieves an accuracy of 99.5% on Labeled Faces in the Wild Dataset [3]. To achieve a similar accuracy on real life scenarios different methods like dimensionality reduction and data augmentation have been implemented. A mobile application has also been developed which utilizes the above described hybrid model for identification of people present in a large group. This application outweighs the fingerprint biometric in terms of speed, cost and group size.