Figure 2

Flow chart of the image reconstruction algorithm. The network input corresponds to the \({\text{j}}\)-th measured diffractive image \(I\) and illuminated position \({{\varvec{r}}}_{{\text{j}}}=({x}_{{\text{j}}}+\delta {x}_{{\text{j}}},{y}_{{\text{j}}}+\delta {y}_{{\text{j}}})\), where \((\delta {x}_{{\text{j}}}\), \(\delta {y}_{{\text{j}}})\) represents the measurement error in the scanning position. The reconstructed complex functions \({O}_{{\text{r}}}\) and \({\psi }_{{\text{r}}}\) are considered as the two-channel learnable filter of the convolutional layer. The number of pixels in \({\psi }_{{\text{r}}} ({m}_{2}\times {m}_{2})\) is determined using FOD, whereas that in \({O}_{{\text{r}}}\) (\({m}_{1}\times {m}_{1})\) is determined using both FOD and FOI. Therefore, a new cropped object function \({O}_{{\text{c}}}\) is defined by selecting \({m}_{2}\times {m}_{2}\) pixels from the illuminated position. Using zero-padding, the pixel number can be adjusted based on the reconstruction time and FOV of the reconstructed image. The output of the network \({I}_{{\text{r}}}\), propagated to the detector plane \(s\) using the element product of \({\psi }_{{\text{r}}}\) and \({O}_{{\text{c}}}\), is compared with the \(I\). A single iteration is completed after the same process is performed for all \(N\) measured images. The optimization process of the network iterates \(i\) times until Eq. (3) is minimized.