Age progression or age synthesis (face aging) is defined as aesthetically rendering a face image with natural aging and rejuvenating effects for a certain face of an individual. This analysis can be used in cross-age face analysis, various authentication system, entertainment, but in finding lost children after a couple of years or more. Researchers from the National University of Singapore have recently developed a method to render aging faces in a personalized way.
The model consists in learning a set of age-group specific dictionaries, where the dictionary bases correspond to the same index, and these form a particular aging process pattern. A linear combination of these patterns express a particular personalized aging process. The personalized part is reflected with taking into account personalized facial characteristics, for example moles, which are invariant in the process of aging. The framework takes into account face pairs from neighboring age groups for a particular subject, since it is almost impossible to collect faces of all age groups for a particular subject.
There are two main categories for the age progression task solution. The first is the prototype-based age progression, which transfers the differences between two prototypes (average faces) of the pre-divided source-age group and target-age group into the input individual face (whose age belongs to the source-age group). The second is the model-based age progression, which models the facial parameters for the shape and texture synthesis with the actual age or age range. To make the aging process more personalized – as it is in real life – the general rules of the aging process have been taken into account, but also the specific process which contains more personalized facial characteristics, such as birthmarks which are almost invariant with time.
Short-term face-aging sequences are available on the Web, such as photographs of celebrities of different ages on social media. Also, some face-aging databases contain dense short-term sequences as well, so generating personalized age progress by leveraging short-term face-aging sequences is more feasible than collecting long-term face-aging sequences, which is very difficult. This method automatically renders aging faces in a personalized way using age-group specific dictionaries. An individual face is decomposed into an aging layer and a personalized layer. The former shows the general aging characteristics, and the later personalized facial ones. The mentioned aging dictionaries are composed to characterize the general human aging patterns, so that the aging layer can be represented by a linear combination of these patterns with a sparse coefficient. The redundancy between the aging layer and the input face can be defined as the personalized layer, invariant in the aging process. The final aging face is rendered by synthesizing the aging layer in the wanted age range and the personalized layer.
To sum up, all aging dictionaries are trained on the collected shirt-term aging database. In two arbitrary neighboring age groups, the younger-age and older-age face pairs of the same people are used to train coupled aging dictionaries with common sparse coefficients, while excluding the personalized layer. For an input face, the personalized aging-face sequence from the current age to the future age is rendered step by step on the learned aging dictionaries. For the future work, the researchers are considering utilizing the bilevel optimization for the personality-aware coupled-dictionary learning model.
At Visage Technologies, we have developed algorithms that can estimate a person’s age in real time based on their unique facial characteristics. To get your free evaluation license, get in touch with us.