Periocular EfficientNet: A Deep Model for Periocular Recognition

Periocular recognition has witnessed significant advancements through computational techniques of pattern recognition, with applications in the fields of security and health. However, this progress is accompanied by challenges such as variations in lighting conditions, image resolution, and partial occlusions. In this study, we propose the Periocular EfficientNet-B0 (PEN) architecture, a customized version of the EfficientNet-B0 convolutional neural network. This architecture is customized through the incorporation of densely connected layers with Softmax for classification purposes. Experimental evaluations conducted on the UFPR-Periocular dataset demonstrate that the proposed model achieves an average accuracy of 99.39%, surpassing the state-of-the-art Linear Discriminant Analysis CNN (LDA-CNN) by 1.71%.