SAM is a neural network model that changes the age of a person in an image. The model takes as input an image of a person’s face and target age. At the output, the neural network gives a generated image, where the target person’s face is changed in accordance with the age shift. SAM can both rejuvenate faces in images and age them. The researchers used the generative adversarial model StyleGAN as the architecture. The model allows you to edit the generated images. Experiments show that SAM bypasses state-of-the-art approaches.
Why is it needed
The task of transforming the age on the image is to change the appearance of a person in accordance with age. Realistic modeling of such a transformation for the input face image requires the model to take into account the change in facial features and head shape and preserve the personality of the person in the photo. The developers propose a neural network that is based on the GAN and is capable of simulating age-related changes in a person’s face.
More about the approach
The researchers solve the problem of simulating the continuous aging process as a regression problem between the input age and the target age. The network receives facial images and target age as input. The encoder first extracts 18 feature maps, which correspond to the 18 StyleGAN style inputs. Then, map2style layers are used, which gradually reduce the dimension of each feature map to one 512 style vector. Researchers additionally use the pre-trained pSp encoder to extract the hidden code, which is then used in the resulting vector of the modified age.
