Computer Vision News - November 2020

Research 12 The results were compared to the state-of-the-art methods adjusted anchored neighbourhood regression A+, FSRCNN, ESPCN, LapSRN, and SRGAN. It is notable that the GAN-CIRCLE predicts images with sharper boundaries and richer textures than GAN-CIRCLE/u which extracts from the learning process additional anatomical information from the unpaired samples. Overall, the proposed methods generate sharper results, comparing to the other methods. The different losses are explained shortly below but read more in the original article, to understand their subtleties. Adversarial loss The adversarial objective with respect of G is: The Wasserstein distance is applied to improve the training quality with gradient penalty. Cycle consistency loss Two main purposes: enforcing latent codes deviating form prior distribution during cycle-reconstruction mapping and prevent the degeneracy during the adversarial learning. Identity loss Which is used to regularise the training procedure. Joint sparsifying transform loss Formulates a non-linear total variation loss to express the image sparsity.