A COMPUTER IMPLEMENTED METHOD FOR TRAINING A GAN NETWORK TO GENERATE ARTIFICIAL IMAGES OF FACES FOR A DRIVER MONITORING SYSTEM

摘要

The invention relates to a computer implemented method for training a GAN network to generate artificial images of faces for a DMS system, said computer implemented method comprising:- configuring a generator of said GAN network to receive an input vector composed of a latent vector and of a labels vector, said labels vector comprising labels extracted from a dataset comprising a primary number of real images of faces of a secondary number of persons,- configuring said generator to generate artificial images of faces, at different scales on each primary convolution layer of which said generator is composed of, based on said input vector,- configuring a discriminator of said GAN network to receive for each secondary convolution layers of which said discriminator is composed of, as inputs an artificial image of a face at a different scale from another secondary convolution layer, and a real image of a face at the same scale, said real image of a face coming from said dataset,- configuring said discriminator to receive said labels vector as an input on an additional secondary convolution layer,- configuring said discriminator to output a loss function,- adjusting weights of said primary convolution layers and weights of said secondary convolution layers based on said loss function using back propagation until a convergence of the generative adversarial network is reached,- applying a metric(s) on said artificial images of faces,- adjusting training parameters to reach a minimum value of the metric(s).