Periocular Region based Gender Identification using Transfer Learning

Abstract

COVID-19 broke out at the end of 2019 and is still affecting the lifestyle of the people. To protect ourselves from the deadly disease, wearing a face mask is recommended when coming in contact with others. The usage of face masks in our daily lives leads to the problem of occlusion for facial image-based gender identification systems. Gender Identification System is an application of computer vision used in biometrics, consumer identification, and security systems. In the situation of masked faces, the only visible part of the face is the area around the eye, i.e., Periocular Region. The motivation behind this research is to build a gender identification system from periocular images with pre-trained CNN models using the Transfer Learning approach. In the proposed methodology, the optimal periocular ROI is first extracted and passed to the different pre-trained CNN models (VGG16, VGG19, ResNet50, ResNet101, Inception V3, DenseNet121) for feature extraction. Then, the fully connected layers are added to the base models for classification. The proposed approach with VGG19, ResNet101, and ResNet50 as the base models outperform existing models with an average accuracy of 98.65%, 98.96%, and 98.99%, respectively, in different experiments on the benchmark UBIPr dataset.

Keywords

• Periocular Region
• Pre-trained CNN Models
• Transfer Learning
• Gender Classification