Abstract
Face recognition is one of the most useful contactless biometric identi cation tool, having
several applications in security, surveillance and human-computer interactions etc. The
concept of face recognition has been familiar in the Computer Vision community since
long, but little development have been achieved due to challenges in feature engineering
and achieving robustness (in terms of pose, illumination, occlusion etc.). However, with
the inception of deep learning, most of these problems have been mitigated. But with all
this improvents, deep learning models require a large amount of data and computational
power to provide an accuracte model.
In this work, we propose a face recognition pipeline based on deep learning architecture
which can be trained with minimum number of samples without hours of training and can
be implemented to run on videos in real time. Our method uses a pretrained deep learning
framework based on the Inception module for feature generation, i.e. this module provides
the feature vector for each face image. These features are used to train a Support Vector
Machine (SVM) for classi cation. This architecture is validated on two large datasets
(LFW and IMFDB).We have tested our model on videos, speci cally movie clips to access
the model accuracy and speed. The achieved accuracy is above 91% with approximately
4-5 frames per second which proves the superority of our model. We have also provided a
real life example of face recognition using our method with very few training images, and
the achieved results are very promising. Finally, we have analysed the factors a ecting
the end result and investigated some of our failure cases for better understanding.