Loss Function of GAN to Make a Clear Judgment

In general, in GAN training, the Discriminator learns a loss function whose value becomes smaller when the expected label matches the correct label. However, if we do not make a clear decision about the authenticity of the image, the GAN will continue to learn about ambiguous images, and the resulting images will be less clear. In this experiment, we propose a loss function that penalizes the Discriminator when it makes an ambiguous decision by adding a regularization term to the existing loss function Binary Cross Entropy(BCE), and verify its superiority through experiments.