Reconstruction

To reconstruct the image, we need the trained epitome and a set of mappings associated with all the patches of the image. The optimal mappings are obtained at the end of the training process, where all posterior probabilities are evaluated on the optimal parameters $e^*$.

$$T_k^* = {\hbox{$\underset{T_k}{\mbox{argmax}}\;$}} f(T_k\vert Z_k, e^*)$$ (14)

In the reconstruction, for each patch in the image, we simply copy its corresponding content in the epitome according to the mapping. As for the pixel covered by multiple patches, we take the mean of all the patches. Some examples of the reconstruction are displayed below, where the color and texture of the original image were well preserved. Note we did not apply smoothing techniques during reconstruction, which accounts for the blocky artifacts in the results.

--> -->

--> -->

--> -->

--> -->

--> -->