Fig. 1

Illustration of the applied CycleGAN architecture for mapping images from domain A to domain B. A real sample image \(X_{A}^{real}\) is mapped to domain B by the generator \(G_{B}: X_{A}^{real} \rightarrow X_{B}^{fake}\) and then back to domain A by the generator \(G_{A}: X_{B}^{fake} \rightarrow X_{A}^{rec}\). The discriminator D_B differentiates between the generated image \(X_{B}^{fake}\) and a real sample image \(X_{B}^{real}\). The same process is done for the reverse direction when mapping a real sample image \(X_{B}^{real}\) from domain B to domain A and backwards, i.e \(X_{B} \xrightarrow {G_{A}} X_{A}^{fake} \xrightarrow {G_{B}} X_{B}^{rec}\). During training, the loss is computed by the adversarial loss \(\mathcal {L}^{adv}\) and the cycle consistency loss \(\mathcal {L}^{cyc}\)