Age Invariant Face Recognition Via Aging Modelling
Aging is a complex problem because at different age points different changes occur in the human face. From childhood to teenage the changes are mostly related to craniofacial growth. At maturity the changes are mostly related to the skin color changes and the face skin starts becoming slack and less smooth. So aging is a mixture of all these components. Moreover, aging is a slow, irreversible and a process that is unique to every human being. Many factors affect the aging process. For example every person has different genes, blood group, and life style and belongs to a particular ethnic group. In order to resolve all these issues we need modelling method that should capture all changes throughout the aging process. So we deal shape and texture separately and finally combine them to develop a synergy between the two. The main objective of thesis is to develop novel aging growth models by using spatiotemporal modeling. It aims to develop and enhance our knowledge of craniofacial growth of human faces and capturing it to form models to understand the changes occurring due to aging process. We investigate five different types of models built on different levels of data granularity and with different perspective. At the global level a model is built on training data that encompasses the entire set of available individuals, whereas at the local level data from homogeneous sub-populations is used and finally at the individual level a personalized model is built for each individual. We used anthropometric features for extracting the shape information of face and for texture information we extract features based on the edges on the texture of the human face. Finally, we make integrated model in order to utilize synergy between three models and arrive at an optimum solution. Our Integrated model is inspired by the Integrated Multi Model Frame Work(IMMF) proposed by Widuputra which exploits synergy that exists between models at the global, local and personalized levels. Our integrated model is based on Adaptive Linear Neuron or later Adaptive Linear Element (Adaline) by calculating weights of each model. A concept of deaging was introduced to align and make our features commensurate with different ages. We built a novel face recognition system based on aging models. It is a two stage process. At first stage we built the model and in second stage we try to identify the probe image into the subspace of the gallery. In the second stage real problem is finding the correct subspace. We have used Naive Bayesian method and the probe image subspace for finding the most probable subspace. The probe image is then searched only in the reduced search space of already determined subspace instead of the whole gallery, thus resulting in reduction of cost of computation time. We built two different sets of models, one for the shape and the other for texture. It was observed that we need to resolve the conflict if our models point to different persons. We enunciated a method to resolve the conflict between two models by building composite models in which we combined the texture and shape. Our composite model is developed by using Decision Tree approach. We have used well known standard databases related to aging, namely, FG-NET and MORPH. Our experimentation showed good results as well as demonstrated utility of our aging models. We also know that aging is primarily a spatiotemporal process. So, tools for spatiotemporal and temporal data (STTD) were also used. The Knowledge Engineering and Discovery Research Institute (KEDRI) has developed methods and tools to deal with STTD, the latest being the NeuCube. We built NeuCube (ST) aging model for age group classification and gender recognition thus demonstrating utilization of a robust tool viz NeuCube (ST).