Deep Gender Classification and Visualization of Near-Infra-Red Periocular-Iris images
Ignacio A. Viedma, Juan E. Tapia · 2018
In this paper, we present an approach of automatic pixels feature extraction for Gender Classification using Near-Infra-Red Periocular iris images with Deep learning. Previous works on gender-from-iris have been tried to find manually the best feature extraction methods to represent the gender information of the iris texture from normalized and encoded images. The application of Soft Biometrics with Deep Learning from NIR Periocular-iris-images is a new topic due to the small number of gender labeled images available. In this work, we used bottleneck, fine-tuning and Convolutional Neural Network (CNN) trained from scratch approaches, to identify the most relevant areas on periocular iris images. Training a CNN from scratch with a small number of images using the Data Augmentation technique reached the best classification rate and automatically found the most relevant areas for this task. We concluded that training a model from scratch even with a small number of layers, performed better than using a pre-trained powerful model such as VGG and Resnet in this kind of problems. The best result reached from our CNN trained from scratch was 85.48% of accuracy for gender classification.