Matija Mlinar (2009) Age classification from facial images. EngD thesis.
Information which we comprehend from faces plays an important role in interaction between humans. If computers would be capable of reliably recognizing information from faces, interaction with computers could be more user friendly. Despite the fact that the age of a person also plays an important role in the interaction between humans, age recognition was not extensively studied so far. For many applications it would be enough if computers could reliably classify facial images into a few age categories such as separating children and adults when accessing the world wide web. In experiments, I made age groups using 5, 10 and 15 year intervals and observed classification accuracy when applying different methods for effective presentation of facial images with less dimensions (PCA, LDA, 2DLDA, statistical shape and appearance models) and when applying different methods for classification of age groups (Gaussian models, polynomial models, nearest neighbours, support vector machines, decision tree). The FG-NET Aging Database where facial images display almost all possible variations was used in experiments. As it was shown the shape of face which is also built into the statistical shape and appearance models has an important advantage over methods which are based only on appearance. Best performance of classification methods was achieved with the SVM method. Using four age groups with 15 year age intervals the highest classification accuracy of 73% was achieved. This indicates that age classification from facial images when using a more naturally defined age groups and more specific facial images could be applied in real applications.
Actions (login required)