MuSTAT: Face Ageing Using Multi-scale Target Age Style Transfer

Most existing bottleneck-based Generative Adversarial Networks suffer from ghosting artifacts or blur for generating ageing results with an increased age gap. Although this can be solved using data collected over long age spans, it is challenging and tedious. This work proposes a multi-scale target age-based style face ageing model using an encoder-decoder architecture to generate high-fidelity face images under ageing. Further, we propose using skip connections with selective transfer units (STU) in the encoder-decoder architecture to adaptively select and modify the encoder feature to enhance face ageing results. Unlike conditional GAN (cGAN) approaches that rely on age as a condition to train the generator, we used style information gathered from a random image of the target age group to train the generator. The qualitative and quantitative results on public datasets show that our model can generate photo-realistic synthetic face images of the target age group. The proposed model can also generate diverse age groups for age progression and regression tasks. Furthermore, the analysis using six different Face Recognition Systems (FRS) also indicates the ability of the proposed approach to preserve the identity resulting in 89.82% False Non Match Rate (FNMR) at False Match Rate (FMR) of 0.0001%.