Image fusion SDDGAN article learning

Semantic-supervised Infrared and Visible Image Fusion via a Dual-discriminator Generative Adversarial Network article study

Model Framework

The network structure of our generator G is shown in the figure.We concatenate the infrared and visible images in the channel dimension and use them as input to g.The output is the final fused image.g consists of five common convolutional layers.For each convolutional layer, the padding is set to the same and the stride is set to 1.Therefore, the size of the feature map will not change.

There are two discriminators Dr and Dv in our network.They have the same architecture.They both function as classifiers, generating scalars to estimate the probability that the input image comes from real data rather than G.

Loss function

Generator loss function

The generator loss LG includes supervision loss Lsup, adversarial loss Ladv, gradient loss Lgrad and MSE loss Lmse.

Supervised Loss Lsup

Against Loss Ladv

Gradient Loss Lgrad

The fourth loss, Lmse, represents the mse loss.We apply the MSE loss to constrain the fused image to contain considerable information from the source image, which can be defined as:

Discriminator loss function