A Deep Learning Based Palmar Vein Recognition: Transfer Learning and Feature Learning Approaches

Abstract

A method of palm dorsal vein recognition using transfer learning and feature learning approaches has been proposed in this paper. This deep learning framework does the learning process automatically so that the features are extracted from the original image without undergoing any preprocessing mechanisms. The system proposed is dealing with two methods of learning: The first method employs the pre-trained models of CNN (AlexNet, VGG19, ResNet50 and ResNet101) for feature extraction from the deeper fully connected (fc6, fc7 and fc8) layers. K-Nearest Neighbour (KNN) and Support Vector Machine (SVM) algorithms are used for classification with combination of Error-Correcting Output Codes (ECOC). The later method employs the transfer learning approach for the extraction and classification of both features with CNN (AlexNet, ResNet50, ResNet101 and VGG19) models. The experiments have been done with Dr. Badawi's dataset of hand veins containing 500 images. In the first method, all the models are given the recognition accuracy that gives promising results when features are extracted from layers of ‘fc6’. Using ECOC with SVM for classification shows greater accuracy rate than the models using ECOC with KNN. Also in models using ECOC with SVM, the ResNet101 model achieves improved performance. The recognition accuracy for all models provides the best result in the experimentation of the second approach when the epoch number is 25 where ResNet101 achieves a 100% recognition rate.