Real time dynamic facial recognition of subject at motion using angular image

Abstract

In the developing world keeping track of violations or implementing a secured environment has become crucial. In order to address such issues dynamic facial recognition could be developed in such a way that it can facilitate and address all these issues. Dynamic facial recognition is a real time recognition of a subject while it is in motion. Different well known pre-trained models for facial recognition such as ResNet50, VGG19, VGG16, DenseNet169, Inceptionv3 and MobileNetv2 were customized according to the requirement of the dataset to bring about the highest accuracy. Before training the models, the process composed of several steps involving data acquisition which retrieved pictures from various angles of subject. To detect faces and create bounding boxes around the faces as well as marking facial landmarks such as eyes, nose and mouth MTCNN algorithm has been used. In order to compare, the test dataset was divided into two different types where one consisted of all the data and the other consisted of only the images with 120 degree deviation. This helped us to understand how feature extraction is an important factor for facial recognition as all the trained models provided improved and better results with the filtered dataset. Among all the models trained, it can be concluded that the best performing model for our custom dataset is VGG19.