Gender classification from periocular NIR images using fusion of CNNs models

Gender classification from periocular images is a challenging topic. Previous algorithms have focused primarily on the use of texture features and not much research has been done on applying Convolutional Neural Networks (CNN) to this task. In this work we trained a small convolutional neural network for the left and right eyes, and more importantly, studied the effect of merging those models and compare it against the model obtained by training a CNN over the fused left-right eye images. We show that the network benefits from this model merging approach, and becomes more robust towards occlusion and low resolution degradation, outperforming the results of using a single CNN model for the left and right set of images. Ex-periments done over a database of near-infrared periocular images show that our CNN model exhibits competitive performance compared to other state-of-the-art methods.