Figure 12.

Deep kernel representation by Li and Wang. ⁷⁸ A U-Net operates on a prior image (e.g. a 3-frame dynamic image series) to output a set of $k$ different feature maps (a feature vector with $k$ elements for each of the $n_p$ pixels). The similarity of feature vectors between pixels is then used to construct the kernel matrix $\mathit{K}$ , subsequently used in standard KEM. Standard KEM estimates a coefficient vector $\mathit{\alpha }$ , such that the reconstructed image $\mathit{x}$ is found by $\mathit{K}\mathit{x}$ . The crucial training of the U-Net is based on mapping a 10x count-reduced reconstruction to a full-count reconstruction of the unique data being reconstructed. DIP, deep image prior; KEM, kernel method; MLEM, maximum likelihood–expectation maximisation.