A Robust Face Recognition model using Deep Transfer Metric Learning built on AlexNet Convolutional Neural Network

摘要

Exertion of Transfer Learning for Convolutional Neural Networks (CNNs) describes a dynamic solution for face recognition instead of building and training a neutral network from scratch. Transfer Learning gains and stores the knowledge while solving one problem, and utilizes it for a different related problem. So, choosing a suitable CNN for Deep Face Recognition among the state of art CNN is a challenging task. Deep face recognition requires a large density of datasets with standard subjects. Subjects with parameters like pose, illumination and expression are major constraints for frontal facial images. Ignoring the parameters in training as well as testing phase will result to significant decrease in recognition rate; also show indelible impact on accuracy. Training the network with improper illumination facial data base can lead to biased classifier in deep CNN. In this paper, we propose AlexNet-CNN architecture-based transfer learning framework to enlarge the facial feature space on the constrained face dataset. The proposed framework is self-possessed with input layer, convolutional layers, activation layers, pooling layers, fully connected layer connected to a classifier layer and runs on standard CUHK (China University of Hong Kong) and ORL (Olivetti Research Laboratory) database in a standard frontal pose faces, under ordinary illumination state and with unbiassed expression ensuing to excellent accuracy in facial recognizing tasks.