Machine Learning for Interpretable Age Estimation

Abstract

Age estimation from facial images has been growing as a machine-learning topic as it has many real-world applications. It can help with security control for minors, human-computer interaction based on age, and law enforcement concerning identity. These various problems could be solved by building and understanding a machine-learning model that labels a facial image into an age range. This comes with its fair share of issues such as people ageing differently due to genetics, the environment, or the facial photo quality. The motivation behind this project is to see which facial features contribute to age. To see this, genetic programming (GP) was specifically used, as its inherent interpretability helps interpret and understand how the model reaches its final age estimation. The results of this project include how accurate the GP is in estimating a person’s age compared to existing solutions, and analysing why the GP picked certain regions over others and how these regions contribute to the final age estimation.