GitHub

10-minute CPU quickstart

This example is intentionally small. It teaches the objects and array shapes without requiring a camera image, GPU, or output files. The synthetic image is only an API demonstration, not a physical hologram generator.

1. Select a backend and make an image

An inline hologram is a 2-D intensity image. ParticleHolography represents it as a real-valued array, normally scaled near the interval 0–1.

using ParticleHolography

backend(:cpu)
n = 32
rows = reshape(Float32.(1:n), n, 1)
cols = reshape(Float32.(1:n), 1, n)
hologram = @. 0.8f0 - 0.15f0 * exp(-((rows - 16.5f0)^2 +
                                     (cols - 16.5f0)^2) / 24f0)
size(hologram)
(32, 32)

2. State the optical geometry

All lengths below use micrometres. Any consistent unit works, but mixing metres and micrometres produces a numerically valid and scientifically wrong result.

wavelength = 0.6328      # μm
pixel_pitch = 10.0       # μm / pixel
front_distance = 800.0   # μm, camera to first reconstructed plane
slice_spacing = 20.0     # μm
slices = 8

grid = propagation_grid(n, wavelength, pixel_pitch)
front = propagation_kernel(-front_distance, wavelength, grid)
step = propagation_kernel(-slice_spacing, wavelength, grid)
nothing

The negative distances back-propagate from the camera toward the object. If your coordinate convention is reversed, reverse the signs consistently.

3. Reconstruct

Gabor reconstruction assumes zero phase at the camera and uses the square root of measured intensity as the wavefront amplitude.

wavefront = gabor_wavefront(hologram)
request = ReconstructionRequest(slices;
    volume=Float32,
    min_projection=N0f8,
)
plan = ReconstructionPlan(front, step; request)
diagnostic = memory_diagnostic(plan, request)
result = reconstruct(plan, wavefront)
(size(result.volume), size(result.min_projection),
 eltype(result.volume), eltype(result.min_projection), diagnostic.safe)
((32, 32, 8), (32, 32), Float32, N0f8, true)

result.volume[:, :, z] is reconstructed intensity at one depth. result.min_projection is the minimum intensity over depth, useful because opaque particles reconstruct as dark regions. The Float32 volume preserves measurement values, while the compact N0f8 projection is convenient for display and storage. Both outputs came from one propagation loop.

4. Reuse the plan for a sequence

FFT plans and the largest work arrays are expensive to create. Build a ReconstructionPlan once for a fixed image shape and geometry, then reuse it.

request = ReconstructionRequest(slices; volume=Float32, min_projection=N0f8)
plan = ReconstructionPlan(front, step; request)
for hologram in frames
    wavefront = gabor_wavefront(hologram)
    result = reconstruct(plan, wavefront)
    # threshold, detect, and save this frame
end

The allocating functions are convenient for notebooks. For tighter control, preallocate arrays and use reconstruct!, xyprojection!, or reconstruct_and_projection!.

Use reconstruct_padded(plan, wavefront; mode=:mean) when the plan was built for a larger working plane and the border should use the input wavefront mean rather than zero.

Next