Fig. 2.

Architecture of the proposed PAT-ADN. PAT-ADN takes unpaired images $u^a$ and $v$ as input ($u^a\in P^a,v\in P$). The image $\hat{u}$ corresponds to the desired output, obtained after $u^a$ effectively removes the artifacts. The encoders consist of $E_P$ and $E_{P^a}=\{E_{P^a}^c,E_{P^a}^a\}$, whereas $G_P$ and $G_{P^a}$ serve as the decoders. The green connection line represents the flow of content encoding to the content decoder, the red connection line represents the flow of artifact encoding to the artifact decoder, and the black connection line represents the flow of other data.