Fig. 2: Network architecture.

a A macro perspective to look at the architecture of the proposed network consisting of two cascaded networks, namely, generation network and elimination network, each of which is composed of an encoder, the latent space, and a decoder. b The interconnection between two networks. For a given input (low frequency), the generation network can generate various candidates (high frequency), and the elimination network will reversely map each candidate into the original space. The optimal candidate is picked out by calculating the Euclidean distance between the input and secondary candidates. c The layer-level illustration of the generation network. The feature extraction module combined with three concatenation layers composes the encoder, while the reconstruction module combined with Concat4 composes the decoder. Sampled variables from the two sets of Gaussian distributions (i.e., variational posterior) constitute the latent space.