Predictive Modeling of Trait-Aging Invariant Face Recognition System Using Machine Learning

Abstract

There is an ever increasing demand for face recognition systems for a wide range of applications. But face recognition systems have a unique challenge, Trait Aging. Trait Aging makes the recognition of individuals’ faces less accurate with time lapse. In this study, a Convolutional Neural Network (CNN) called Inception-Resnet-V2 was adapted using transfer learning to develop a trait aging invariant face recognition system. The adapted neural network model was tested and trained using image datasets from the MORPH database. The MORPH dataset used in this study had 55,000 unrepeated images of over 13,000 persons. The images were acquired between the year 2003 and 2007. The median age of the dataset was 33, while the minimum and maximum ages in the dataset were 16 and 77 respectively. The mean separation in days between the time each image in the dataset was acquired was 164 days with 1 day and 1681 days as the minimum and maximum separation in days respectively. The standard deviation of separation in days for all images in the dataset was 180 days. Training and testing accuracies reached as high as 100% and experimental results showed training and testing loss of as low as 0%. All these, notwithstanding the trait aging factor present in the pre-processed MORPH database.