We propose a short-path U-net model with average pooling in order to restore in-focus images of multiple transparent particles at ground truth z positions and simultaneously remove their zero-order images, conjugate images, the defocused images of the other particles, and the noise induced by the optical system in a volume. Subsequently, we obtain the lateral location and diameter of each particle in a fast way based on the nature of the particles’ shape; eventually acquire the particle density in the captured volume. The experimental results demonstrate that all sized particles including small– sized and relatively large-sized particles can be restored well by a short-path U-net model, and average pooling performs well when dealing with piece-wise pattern. The proposed scheme can distinguish as many particles as possible in a small volume with small reconstruction depth spacing, such as 50 μm.
|