[ 1 ] [ 2 ] [ 3 ] [ 4 ] [ 5 ] [ 6 ] [ 7 ] [ 8 ] [ 9 ] [ 10 ] [ 11 ] [ 12 ] [ 13 ] [ 14 ]

Journal of Information Science and Engineering, Vol. 37 No. 4, pp. 779-792

Successive Multitask GAN for Age Progression and Regression

Department of Mechanical Engineering
National Taiwan University of Science and Technology
Taipei, 10607 Taiwan
E-mail: fbirken1018;anson183785g@gmail.com, fjisong@mail.ntust.edu.tw

Due to recent progresses made by state-of-the-art deep learning approaches, the facial age progression and regression has become an attractive research topic in the fields of computer vision. Many existing approaches require paired data which refer to the face images of the same person at different ages. As the cost of collecting such paired datasets is expensive, some emerging approaches have been proposed to learn the facial age manifold from unpaired data. However, the images generated by these approaches suffer from the weakness in generating some age traits, for example wrinkles and creases. To generate better age traits, we propose the Successive Multitask GAN (SM-GAN) for age progression and regression. The SM-GAN consists of n triple networks, [T0, T1, ..., Tn-1], and a face feature extractor C. Each triple network Ti consists of a generator Gi, a discriminator Di and a multitask classifier Mi, i.e., Ti = [Gi, Di, Mi]. Gi is trained for transforming between neighboring age groups. Di is trained to distinguish the generated faces from the real faces in each age group in the training set. Mi is trained for age and gender classification. The face feature extractor C warrants the identity consistency between the input and the generated output of Gi. The pixel-wise loss is also exploited to maintain the image attributes between the input and the generated output.To better define the age groups appropriate for successive age generation, we propose a facial age clustering approach to better determine the boundary ages needed for age segmentation. Experiments show that the proposed SM-GAN can generates better facial age images with better age traits compared with other contemporary approaches.

Keywords: generative adversarial network, face generation, facial age transformation, age progression and regression, face recognition

  Retrieve PDF document (JISE_202104_03.pdf)