Age Estimation in Facial Images

Author:

Pisupati, Venkata Soumya

Graduate Program:

Computer Science and Engineering

Degree:

Master of Science

Document Type:

Master Thesis

Date of Defense:

September 26, 2013

Committee Members:

Robert T Collins, Thesis Advisor/Co-Advisor

Keywords:

Age Estimation
Hierarchical estimation
Gaussian Processes
EM algorithm
Image registration
Eye coordinate localization

Abstract:

Automatic age estimation in human facial images is important because people’s behavioral patterns, lifestyles and preferences vary along with their age, and preferred modes of human-computer interaction are different for people of different age groups. An approach for automatic age estimation from facial images therefore has the potential for a number of practical applications. Human beings develop the ability to accurately estimate age early in life, but automatic age estimation is very sensitive to factors such as race, gender, facial attachments such as glasses, beards, facial piercings, and other visually perceptive factors like attractiveness, thus making it one of the challenging problems in computer vision. In this thesis we develop computer vision and machine learning based approaches for extracting visual features from a human facial image and mapping those to an output label estimating the approximate age (in years) or age group (range of years) of the individual. We develop an automatic eye localization algorithm which finds the coordinates of both the eye centers. These co-ordinates are used to register all the facial images into a common frame of reference. Such a localization algorithm is very useful in real world applications where information about eye-center is usually not available. We then extract feature descriptors that represent the facial image in a low dimensional space. In this thesis we have experimented with three different kinds of age relevant feature descriptors in order to capture the shape and texture information of facial images: Discrete Cosine Transform (DCT), Local Binary Pattern (LBP) and Gradient Orientation Pyramids (GOP). We then propose two machine learning pipelines for age estimation. First, we develop a hierarchical approach to estimate the age of the test image using two steps. In the first step we perform soft classification by separating the training data into age groups and using a kNN/SVM classifier to compute the probability of a new test image belonging to each of the age three groups. We then use Gaussian Process (GP) regression for the final age estimation. Gaussian process regression is a Bayesian approach that computes the posterior density for the age of a test image given both the training and test data. GP produces confidence intervals for the age, thereby providing the user with more information. Second, we develop expectation maximization (EM) framework to jointly address the issues of categorizing the training data points into groups and learning the hyper parameters of the Gaussian Process regression model corresponding to each group. Unlike the earlier approaches where we manually categorize the training data into three groups based on age, this approach automatically groups the data. We test our methods on the publicly available FGNET aging database, a very popular and challenging database for testing facial aging based algorithms. We evaluate our algorithms using the Leave One Person Out (LOPO) evaluation scheme. In LOPO we train the regressor on images of all the persons in the database except one, and test it on the images of that one person which have not been used in the training. We repeat this for all the people in our database and compute the error by averaging across errors from the individual testing rounds. This metric is known as the Mean Absolute Error (MAE) and has been used as the performance metric to evaluate performance of various features and our estimation algorithms.