Single-sample face and ear recognition (SSFER) is a challenging sub-problem in biometric recognition that refers to the difficulty in feature extraction and classification when only a single-face or ear training image is available. SSFER becomes much more challenging when images contain a variety of lighting, positions, occlusions, expressions, etc. Virtual sample generation methods in SSFER have gained popularity among researchers due to their simplicity in the augmentation of training sets and improved feature extraction. In this article, we propose a novel and simple method for the generation of virtual samples for training the classifiers to be used in SSFER. The proposed method is based on 2D local patches, and six training samples are generated for a single face or ear image. Further, training is performed using one of the variations along with its generated virtual samples, while during testing, all the variations were considered except the one used during training. Features are extracted using principal component analysis, and classification is performed using the nearest-neighbour classifier. Extensive experiments were performed for the image quality of the virtual samples, classification accuracy, and testing time on ORL, Yale, and AR (illumination) face databases, and AMI and IITD ear databases which are publicly available. The results are also compared with other state-of-the-art methods, with classification accuracy and universal image quality being the major outcomes. The proposed method improves the classification accuracy by 14.50%, 1.11%, 0.09%, 21.60%, and 10.00% on AR (illumination), Yale, ORL, IITD, and AMI databases, respectively. The proposed method showed an improvement in universal image quality by 15%, 20%, 14%, 30%, and 15% on AR (illumination), Yale, ORL, IITD, and AMI databases, respectively. Experimental results prove the effectiveness of the proposed method in generating virtual samples for SSFER.
