Research Area:  Machine Learning
Deep learning methods, especially convolutional neural networks have achieved significant success in the area of computer vision including the difficult face recognition problems. Training of deep models shows exceptional performance with large datasets, but they are not suitable for learning from few samples. This paper proposes a modified deep learning neural network to learn face representation from a smaller dataset. The proposed network is composed of a set of elaborately designed CNNs, RELUs and fully connected layers. The training dataset is augmented with synthetically generated samples by applying Gaussian and Poisson noise to each sample of the training set, thus doubling the size of the training set. We experimentally demonstrate that the augmented training dataset actually improves the generalization power of CNNs. The network is trained using the standard AT&T face database. Using the proposed approach for limited training data, substantial improvement in recognition rate is achieved.
Keywords:  
Author(s) Name:   Umme Aiman; Virendra P. Vishwakarma
Journal name:  
Conferrence name:  8th International Conference on Computing, Communication and Networking Technologies (ICCCNT)
Publisher name:  IEEE
DOI:  10.1109/ICCCNT.2017.8203981
Volume Information:  
Paper Link:   https://ieeexplore.ieee.org/abstract/document/8203981