Improving Gender Classification based on Skin Color using CNN Transfer Learning

Abstract

Gender is one of the important features in soft biometrics. Gender classification has its own importance that can further improve facial recognition performance. However, there is an issue related to inaccurate gender classification from facial images especially for dark-skinned women. In this paper, a gender classification study was conducted to investigate face image classification based on skin color. Transfer learning method is used to improve the accuracy of gender classification. The MobileNet model was selected in this research. Two types of transfer learning namely transfer learning through feature extraction and transfer learning through fine tuning have been implemented. Performance is measured by considering the accuracy of bright and dark skin faces. The experimental results show that, with proper training techniques, gender classification can achieve competitive classification accuracy by using transfer learning methods.