Age and gender estimation using optimised deep networks

Age and gender estimation plays a fundamental role in intelligent applications such as access control, marketing intelligence, human-computer interaction etc. The advent of deep architectures have paved a way to improve the performance of estimation models, however, there is still a lack of optimized architectures. This paper focuses on the use of convolutional neural networks, and parameter modeling and optimization, and their effect on accuracy and loss term convergence. This paper first makes use of a generalized deep architecture based on literature and looks at ways of optimizing and reducing complexity without loss of accuracy. Different activation functions such as rectified linear unit (ReLU), linear function, exponential linear unit (ELU), hyperbolic tangent and Googles' proposed Swish function were tested along with the use of additional convolutional and fully-connected layers. Experiments resulted in a less complex architecture for gender classification and results were in line with that of benchmark accuracies found in literature, however, the same couldn't be achieved for age estimation. The inability to find a simpler architecture for age estimation is attributed to the complex nature of features that are associated with age than that of gender and also the multi-class classification nature of the age estimation problem.