Post-Processing Networks

A post-processing network refines the matched filter output:

$$\hat{\mathbf{x}} = f_\theta(\mathbf{A}^H\mathbf{y})$$

Typically a U-Net trained on (input, ground truth) pairs.

Unrolled algorithms replace fixed iteration parameters with learnable ones:

$$\mathbf{x}^{(k+1)} = \mathcal{S}_{\lambda_k}\left(\mathbf{x}^{(k)} + \frac{1}{L_k}\mathbf{A}^H(\mathbf{y}-\mathbf{A}\mathbf{x}^{(k)})\right)$$

where $\{\lambda_k, L_k\}$ are learned per iteration.