Deep Convolutional Neural Network-Based Recognition of Profile Rotated Face Images

Over the past few decades, various face recognition algorithms have gained widespread deployment in different biometric access mechanisms. While earlier face recognition techniques utilized hand-picked features for computing the distance metric between face images, the advent of deep learning based techniques have automated this process and has revolutionized face recognition technology. However, most of the existing face recognition techniques are trained on frontal face posture and have very poor accuracy when there is a change in facial posture. Moreover, they have poor recognition rate when training dataset is small. In this article, we propose a dataset augmentation based approach for face recognition with convolutional neural networks (CNN). Proposed approach is especially suited for the datasets which consist of limited number of profile rotated facial images with varying brightness. Through extensive simulations, superiority of proposed approach is established by comparing the recognition rate with some of the recently proposed techniques in literature.