Infrared-Based Gender Recognition from Facial Images: A Comparative Study Using Deep Learning

The gender is an important soft biometric trait of a person, upon which many social and economic functionalities depend or rely. Automatic recognition of the gender of individuals is crucial to many real-world applications, such as public video surveillance, gender-specific access and services, human-computer interaction, demographical studies. Up until now, most research works on gender recognition focus on the visible light, leaving the problem of infrared-based gender recognition scarcely studied. Thus, this research work studies IR facial images-based gender recognition, using deep learning models. We design a CNN model on top of ResNet using a multiscale technique (named as IR-GenderNet). Then, we compare the performance of gender recognition between the visible light and IR bands. We also compare the performance of IR-GenderNet against other methods in the literature that are either traditional or deep learning-based. Experimental results demonstrate that gender recognition using IR is not just feasible but even more beneficial than visible light. It is also proven that the proposed model of IR-GenderNet performs better than all other methods for the problem of IR-based gender recognition. Lastly, cross-dataset experiments also justify the generalization ability of IR-based gender recognition using the proposed algorithm.