Two-Level Image Augmentation with Deep-Learning for Single Sample Face and Ear Recognition

Single-sample face and ear recognition is a challenging sub-problem in biometric recognition, as only one sample is available for training. This article presents a method to recognise the subject from a single face or ear image with two-level image augmentation, followed by using a pre-trained deep learning model for feature extraction and classification. First, sampling is used for generating additional samples, and then mirroring and rotation are used. Pre-trained VGG16 model is used for parameter estimation and classification. The proposed method is tested on three face and two ear databases. The proposed method achieves classification accuracy of 84.83 %, 77.08 %, and 88.52 % on ORL, Yale, and FeedTum databases and 66.69% and 78.29% on AMI and IITD databases, respectively. The image quality of the generated samples is also found to be in an acceptable range. The proposed method can be helpful and effective when smaller databases and fewer training images are available.