The Computer Nose Best

The nose is the most central feature on the face which is known to exhibit both gender and ethnic differences. It is a robust feature, invariant to expression and known to contain depth information. In this paper we address the topic of binary ethnicity classificiation from images of the nose, using a novel dataset of South Asian, Pakistani images. To the best of our knowledge, we are one of the first to attempt demographic (ethnicity) based identification based solely on information from the nose. A two-category (Pakistani vs Non-Pakistani) task was used in combination with Deep learning (ResNet) based and VGG-based pre-trained models. A series of experiments were conducted using ResNet-50, ResNet-101, ResNet-152, VGG-Face, VGG-16 and VGG-19, for feature extraction and a Linear Support Vector Machine for classification. The experimental results demonstrate ResNet-50 achieves the highest performance accuracy of 94.1%. In comparison, the highest score for the VGG-based models (VGG-16) was 90.8%. These results demonstrate that information from the nose is sufficient for deep learning models to achieve >90% accuracy on judgements of ethnicity.