Face analysis and age estimation

Abstract

Facial images convey many significant human characteristics, identity, gender, expression and so on. While face and gender recognition have been extensively studied in the computer vision and pattern recognition community, facial age estimation has not been well addressed. To explore computational facial age estimation, this project was carried out in the following three phases. In the first phase, previous researches on facial aging patterns, facial features, and age estimation methods were studied. Then two principal facial aging features, shape and texture for 1002 faces, 82 people ranging from 0 years old to 68, were extracted and organized in a pre-defined manner, preparing for the next step analysis. In the second phase, with all the features obtained, Principle Component Analysis (PCA) and Linear Discriminant Analysis (LDA) were applied to examine the efficacy of the extracted features, respectively. For the texture features, various color spaces were also explored and the best one is picked out. Most existing facial age estimation methods, however, usually unitize only the appearance features (texture information) of facial images for age estimation. Since shape information also contributes to age estimation, this project was motivated to examine the combination of multiple features for age estimation. Thus in the final part, a new method was proposed. This algorithm utilized Canonical Correlation Analysis (CCA) to fuse both the shape and texture information of facial images to characterize human ages. To uncover the relation of the fused features and the ground-truth age values, a multiple linear regression function with a quadratic model was learnt for age estimation. The new method demonstrated significant improvement in facial age estimation performance.