Loss Function

The discriminator model can be updated directly, whereas the generator model must be updated via the discriminator model. This can be achieved by defining a new composite model that uses the output of the generator model as input to the discriminator model. This composite model involves stacking the generator on top of the discriminator. The generator is updated to minimize the loss predicted by the discriminator for the generated images marked as “original.” As such, it is encouraged to generate more realistic images. The generator is also updated to minimize L1 loss or mean absolute error between the generated image and the target image. This is accomplished by using the weighted sum of both the adversarial loss from the discriminator output and the L1 loss (100 to 1 in favor of L1 loss) to update the generator. This weighing encourages the generator strongly towards generating more realistic translations of input images in the target domain. We implemented the model architecture and configuration proposed by Isola et al.,12 with minor modifications needed to generate color images of 512 × 512 pixels using Keras 2.3.1 and Tensorflow 2.0.0 libraries.21^,22 The proposed model architecture is illustrated in Figure 2.

A simplified plot of the proposed composite pix2pix model outlining its main components and workflow (created by Hazem Abdelmotaal). 1: generated image. 2: discriminator loss (D = 0.5 × discriminator cross-entropy loss). 3: adversarial loss. 4: composite loss function (generator loss = adversarial loss + lambda (10) × L1 loss). L1 loss = mean absolute error between the generated image and the target image.