Skip to content

Optimizing holograms using Odak

This engineering note will give you an idea about how to optimize phase-only holograms using Odak. We consult the beginners in this matter to Goodman's Introduction to Fourier Optics (ISBN-13: 978-0974707723) and Principles of optics: electromagnetic theory of propagation, interference and diffraction of light from Max Born and Emil Wolf (ISBN 0-08-26482-4). Note that the creators of this documentation are from the Computational Displays domain. However, the provided submodules can potentially aid other lines of research as well, such as Computational Imaging or Computational Microscopy.

The optimization that is referred to in this document is the one that generates a phase-only hologram that can reconstruct a target image. There are multiple ways in the literature to optimize a phase-only hologram for a single plane, and these include:

Gerchberg-Saxton and Yang-Yu algorithms: - Yang, G. Z., Dong, B. Z., Gu, B. Y., Zhuang, J. Y., & Ersoy, O. K. (1994). Gerchberg–Saxton and Yang–Gu algorithms for phase retrieval in a nonunitary transform system: a comparison. Applied optics, 33(2), 209-218.

Stochastic Gradient Descent based optimization: - Chen, Y., Chi, Y., Fan, J., & Ma, C. (2019). Gradient descent with random initialization: Fast global convergence for nonconvex phase retrieval. Mathematical Programming, 176(1), 5-37.

Odak provides functions to optimize phase-only holograms using Gerchberg-Saxton algorithm or the Stochastic Gradient Descent based approach. The relevant functions here are odak.learn.wave.stochastic_gradient_descent and odak.learn.wave.gerchberg_saxton. We will review both of these definitions in this document. But first, let's get prepared.


We first start with imports, here is all you need:

from odak.learn.wave import stochastic_gradient_descent, calculate_amplitude, calculate_phase
import torch

We will also be needing some variables that defines the wavelength of light that we work with:

wavelength = 0.000000532

Pixel pitch and resolution of the phase-only hologram or a phase-only spatial light modulator that we are simulating:

dx = 0.0000064
resolution = [1080, 1920]

Define the distance that the light will travel from optimized hologram.

distance = 0.15

We have to set a target image. You can either load a sample image here or paint a white rectangle on a white background like in this example.

target = torch.zeros(resolution[0],resolution[1])
target[500:600,400:450] = 1.

Surely, we also have to set the number of iterations and learning rate for our optimizations. If you want the GPU support, you also have to set the cuda as True. Propagation type has to be defined as well. In this example, we will use transfer function Fresnel approach. For more on propagation types, curious readers can consult Computational Fourier Optics David Vuelz (ISBN13:9780819482044).

iteration_number = 100
learning_rate = 0.1
cuda = True
propagation_type = 'TR Fresnel'

This step concludes our preparations. Let's dive into optimizing our phase-only holograms. Depending on your choice, you can either optimize using Gerchberg-Saxton approach or the Stochastic Gradient Descent approach. This document will only show you Stochastic Gradient Descent approach as it is the state of art. However, optimizing a phase-only hologram is as importing:

from odak.learn.wave import gerchberg_saxton

and almost as easy as replacing stochastic_gradient_descent with gerchberg_saxton in the upcoming described hologram routine. For greater details, consult to documentation of odak.learn.wave.gerchberg_saxton.

Stochastic Gradient Descent approach

We have prepared a function for you to avoid compiling a differentiable hologram optimizer from scratch.

hologram, reconstructed = stochastic_gradient_descent(
        'TR Fresnel',
Iteration: 99 loss:0.0003

Congratulations! You have just optimized a phase-only hologram that reconstruct your target image at the target depth.

Surely, you want to see what kind of image is being reconstructed with this newly optimized hologram. You can save the outcome to an image file easily. Odak provides tools to save and load images. First, you have to import:

from import save_image,load_image

As you can recall, we have created a target image earlier that is normalized between zero and one. The same is true for our result, reconstructed. Therefore, we have to save it correctly by taking that into account. Note that reconstructed is the complex field generated by our optimized hologram variable. So, we need to save the reconstructed intensity as humans and cameras capture intensity but not a complex field with phase and amplitude.

reconstructed_intensity = calculate_amplitude(reconstructed)**2

To save our hologram as an image so that we can load it to a spatial light modulator, we have to normalize it between zero and 255 (dynamic range of a typical image on a computer).

P.S. Depending on your SLM's calibration and dynamic range things may vary.

slm_range = 2*3.14
dynamic_range = 255
phase_hologram = calculate_phase(hologram)
phase_only_hologram = (phase_hologram%slm_range)/(slm_range)*dynamic_range

It is now time for saving our hologram:


In some cases, you may want to add a grating term to your hologram as you will display it on a spatial light modulator. There are various reasons for that, but the most obvious is getting rid of zeroth-order reflections that are not modulated by your hologram. In case you need it is as simple as below:

from odak.learn.wave import linear_grating
grating = linear_grating(resolution[0],resolution[1],axis='y').to(phase_hologram.device)
phase_only_hologram_w_grating = phase_hologram+calculate_phase(grating)

And let's save what we got from this step:

phase_only_hologram_w_grating = (phase_only_hologram_w_grating%slm_range)/(slm_range)*dynamic_range

See also

For more engineering notes, follow: