Radio Galaxy Morphology Simulation via Residual Conditional Variational Autoencoder

We propose a radio galaxy morphology simulation approach by using a conditional variational autoencoder composed of residual convolutional blocks namely ResCVAE. It estimates the distribution of the existed morphology-labeled radio galaxy images by mapping them to a latent low-dimension space, and simulates new images according to the obtained distribution. From designed experiments we find that the pixelwise cross-entropy (PCE) loss, as a component of the optimization objective for the ResCVAE, trains the network parameters to obtain lower reconstruction loss and more accurate simulation performance than the mean squared error (MSE) loss function under the same training settings. In addition, the strategies of applying condition and residual convolutional blocks are also evaluated and compared to the other networks, which suggests a domination of our proposed approach.