Combining regression and classification methods for age band estimation from human faces

This work is made under a behavioral marketing project, and aims at identify the age band of a person from a 2D image face. A Multi-Scale and Multilevel Histogram of Blocks of Local Binary Patterns (MMHLBP) is used as texture descriptor. Three approaches are performed and compared. In the first approach, the goal is to predict the age before getting the age band. Partial Least Squares Regression and Support Vector Regression are implemented. In the second approach, the problem is solved as a classification problem. A combination of Support Vector Machine and Classification and Regression Tree is performed. The results show that second approach overcomes the first approach; but the second approach generates extreme confusions. The regression method and the classification method are combined in a third approach, which leads to better results. Eventually, a comparison between human operator and our final model shows the relevancy of the combination between the regression approach and the classification approach.