Figure 1
The `double DIP' model used in this work. Building on prior work (Gandelsman et al., 2018), the outputs of both image-generating deep image priors (DIPs; as shown in Fig. 2) are filtered by function f1/2 to account for the partial band transfer of superimposed images. f1/2 can be either a single-layer filter, or a shallow three-level DIP network. |