AG Kommunikationstheorie

Thema:

Face Image Description by Boosted Classification and Regression

Abstract:

Recently, face image description has been applied in more and more fields like electronic passport or ID, public security, criminal investigation and biomedicine. Thus, our project is focused on this topic by using boosted methods. In the past years, most of the researches are concentrated on the classification for the face image description, while in this project we are attempting to apply the boosting method not only to the face image classification but also to the face image classification regression. Considering the popularity of face image description at present, we think it is of great significance to look into this subject.

In this project, we investigate three topics: to classify the gender of a person; to estimate the age of a person and to predict the left eye position of a person. For the first part, the Adaboosting method is introduced to decide the gender. For the second and third part, the gradient tree boosting is applied to estimating the continuous variables: age and left eye position.

The results of the experiments are acceptable. The best classification error rate of gender is around 15%. The correct rate of left eye position is about 88%. But if we set the tolerance of error as one pixel for each axis, it becomes 99%.While the average error for estimating the age is nearly 8 years.



zurück zur Terminübersicht