Model Agnostic Few Shot Learning for Finger Vein Recognition

In the field of biometrics and medical imaging, research on finger vein biometrics has been active and effective with limited datasets that are publicly available to perform experimentation. However, with how secure and reliable the finger vein images are used and with limited set of examples or users with limited samples there is a high need to create solutions to tackle the classification of limited users which would be used for limited accesses. In this paper we have focused on building various ensemble techniques along with few shots learning methodology which is model agnostic and would give the best accuracy in any possible scenario by using convolutional neural network (CNN) and Siamese network along with usage of the weights of couple of pretrained image models. Three different approaches with fusion mechanism has been proposed to predict better even with very few examples of the data. The various performance metrics, hyperparameter tuning, feature extraction, tables and figures are explained in the paper.