SCATTERING MEDIA

BACKGROUND OF THE INVENTION

1. FIELD OF THE INVENTION

The present invention relates to a system for guiding light through scattering media with the aim of projecting perceivable images of objects and natural scenes on a screen placed after the scattering medium. The projection is achieved with an inten sity only detection. No interferometric detection is needed, which is a significant improvement over prior art. Despite this partial measurement, the fidelity of the projected images is on a par with that of the gold standard full measurement (intensity and phase) systems.

2. BACKGROUND ART

Transmission of light through a scattering medium is a challeng ing task. The propagating light entering such a medium could un- dergo strong distortion depending on the medium, for example but not limited to a multimode fiber (MMF) , because of energy split ting of thousands of the modes being excited in the medium.

Several approaches have been proposed to compensate the modal dispersion in MMFs . Analog phase conjugation via holographic ma terials initially suggested by Wertz and Spitz in 1966 (Spitz, E. & Werts, A. Transmission des images a travers une fibre optique. Comptes Rendus Hebd. Des. Seances De. L Acad. Des. Sci. Ser. B 264, 1015, 1967.) and later on via a third order nonline ar crystal by Yariv (Gover, A., Lee, C. P. & Yariv, A. Direct transmission of pictorial information in multimode optical fi- bers . JOSA 66, 306-311, 1976.) was the first method proposed for recovering the distorted fields. Many decades later, digital ho lography and phase conjugation was carried out digitally on the computer (Yamaguchi, I. & Zhang, T. Phase-shifting digital ho lography. Optics letters 22, 1268-1270, 1997./ Papadopoulos , I. N., Farahi, S., Moser, C. & Psaltis, D. Focusing and scanning light through a multimode optical fiber using digital phase con jugation. Optics express 20, 10583-10590, 2012.) . Nevertheless, today it still requires an extremely accurate calibration and is significantly prone to perturbation.

Alternatively, digital iterative algorithms implemented with spatial light modulators (SLMs) are used to shape the beam en tering the MMF to form a focused spot at the output of the fiber (Di Leonardo, R. & Bianchi, S. Hologram transmission through multi-mode optical fibers. Optics express 19, 247-254 2011./

Cizmar, T. & Dholakia, K. Shaping the light transmission through a multimode optical fibre: complex transformation analysis and applications in biophotonics. Optics Express 19, 18871-18884,

2011) . By scanning the focused spot in the field of view of the fiber, an image can be produced. Other methods use interferometry to measure the complex field at the output of the fiber and then construct a transmission matrix between the light modulator component (SLM for example) and the object on both sides of the MMF (Choi, Y. et al . Scanner-free and wide-field endoscopic imaging by using a single multimode optical fiber. Physical review letters 109, 203901, 2012./ Cara- vaca-Aguirre, A. M., Niv, E., Conkey, D. B. & Piestun, R. Real time resilient focusing through a bending multimode fiber. Op tics express 21, 12881-12887, 2013./ Gu, R. Y., Mahalati, R. N. & Kahn, J. M. Design of flexible multi-mode fiber endoscope. Op tics express 23, 26905-26918, 2015) . This method entails measur ing the phase with holography. Additionally, after any small perturbation in the system the matrix should be re-measured, which could be a challenging task in terms of time and computa- tional resources. A machine learning approach which utilizes conventional simplistic training procedures is also used to fo cus spots through MMFs . The ability of this approach in projec tion of images is limited to focusing spots or simple shapes without the ability to generalize to other classes never seen by the network (Turpin, A., Vishniakou, I. & d Seelig, J. Light scattering control in transmission and reflection with neural networks. Optics express 26, 30911-30929, 2018) .

It should be noted that out of the two major approaches for light transmission inside the MMF, i.e. the transmission matrix and the digital iterative method, the former can project any ar bitrary pattern at the output of the fiber but requires nontriv ial experimental setup, and the latter is limited to projecting simple and basic shapes such as focused spots.

Hence there is a need for a system that while being able to pro ject a wide range of images through the scattering medium, is simple enough to implement. SUMMARY OF THE INVENTION

The present invention circumvents all of the previous shortcom ings of guiding light through the scattering media.

In detail, the present invention is related to a method for pro- jecting images through or by reflection off a scattering medium, the method comprising the steps of:

Illuminating a light modulator, preferably selected from the group consisting of a spatial light modulator (SLM) and a digital micromirror device (DLD) , of a multimode optical fiber projector with one or more light source (s);

Providing a modulation pattern selected from the group consisting of amplitude-only modulation, phase-only modu lation, amplitude and phase-modulation, to the light mod ulator to modulate spatially the light from said light source ; Sending said modulated light through the scattering medi um or by reflection off the scattering medium onto a screen;

Detecting the amplitude of the modulated light with a camera system; and

computing a spatial modulation pattern by using a system of two or more neural networks in series.

The present invention is furthermore related to a multimode op tical fiber projector for performing the method of the present invention, comprising

One or more light source (s), preferably a laser source or an LED,

A light modulator, preferably selected from the group consisting of a spatial light modulator (SLM) and a digi tal micromirror device (DLD) ,

a device comprising a scattering medium, such as a multi- mode fiber (MMF) ,

a camera system comprising a screen, and

a system of two or more neural networks in series for computing a spatial modulation pattern. Detailed description

A first embodiment of the present invention comprises a training of one of the neural networks in a path of propagation of light in forward direction, i.e. from the light modulator such as a spatial light modulator (SLM) or Digital Micromirror Device (DMD), but not limited to, until the screen, i.e. an intensity detector, through the scattering medium. The detection with a CMOS camera or a CCD camera is performed non- interferometrically, meaning that only the amplitude is detect ed .

This can be done by collecting examples displayed on the light modulator such as the SLM or DMD, which could modulate a light beam incident on it by amplitude-only modulation, phase-only modulation or a combination of both, and collecting the corre sponding output fields on the screen. The digital neural network (in the computer) then could be trained with these experimental ly collected dataset.

Another embodiment of the present invention comprises a training of one of the neural networks in a reverse path, i.e. from the screen to the light modulator, for example a SLM, through propa gation inside the scattering medium. This screen and SLM are re- ferred to as the original screen and original SLM. Then, the training of the neural network in the reverse path can be done in several ways: As an example, but not limited to, by adding an extra screen and extra light modulator, for example SLM, to the setup, so that in total the setup consists of two SLMs and two screens. The extra SLM added to the system must be placed in a position with the same optical path difference as the original screen. Similarly, the extra added screen must be placed in a position with the same optical path difference as the original

SLM. In this way, by generating SLM patterns in the extra added SLM device and capturing the output patterns on the extra added screen, and then training the neural network based on these pairs of inputs-outputs from the extra SLM and extra screen, the reverse path of the original setup is learned. So, when the goal is to project a perceivable image on the original screen, a training set consisting of images from, but not limited to, the same class are first uploaded to the extra SLM, and their corre sponding outputs transmitted through the fiber in the reverse path are captured on the extra screen. By training a neural net work with these two sets of input-output images, it is possible to project perceivable patterns in the forward way to be cap tured on the original screen.

In another embodiment according to the present invention, the training in the reverse optical path could be carried out by us- ing an implementation akin to Generative Adversarial Networks (GANs) . In more detail, two sets of neural networks are used in this embodiment. One set of neural networks (designated here as set of networks D) is the one that emulates the forward path of light propagation through said scattering medium. The second set of neural networks (designated here as set of networks G) learns the reverse optical path of light propagation through said scat tering medium. The training of the first set (i.e. neural net works emulating the forward path) is performed as explained in the first embodiment. The second set of neural networks is trained synergistically with the first set so that the discrimi nator (i.e. the first set of neural networks) forces the genera tor (i.e. the second set of neural networks) to output SLM pat terns that, upon sending through the scattering medium (e.g. a fiber), produce images on the screen similar to the desired set. The training of the two networks, depicted schematically in Fig.

3b, may be carried out in the following way: initially neither of the networks is trained; hence the training may start by col lecting examples for training network D. These examples may con sist of SLM images and their corresponding output speckle pat- terns on the screen. It should be noted that the goal here is to find a subset domain of SLM images that produce a certain class of images on the output of the scattering medium (e.g. a fiber) . In the beginning, there is no information about such a domain.

Hence, in a first iteration, random SLM images are sent through the scattering medium (e.g. a fiber) and the images are measured on the screen so as to produce a dataset, and then the first set of networks D is trained with this dataset. Once training of the first set of networks D is complete, training of the second set of networks G begins by feeding it with a class of images that should be projected through the scattering medium (e.g. a fi ber) . The second set of networks G learns the mapping from the output amplitude images on the screen to the proper SLM images constrained with the physical propagation rules enforced by the first set of networks D. By the end of the first iteration, the domain of the output SLM patterns generated by the second set of networks G must have moved closer in some metrics (Euclidean distance, for example) to the domain of SLM images that produce perfect desired images at the output of the scattering medium (e.g. a fiber) .

In a second iteration, training of the first set of networks D is carried out using the SLM images that were produced by the second set of networks G in the previous (first) iteration ra ther than random images as in the first iteration. Once com- plete, training of the second set of networks G is carried out in the same way as in the first iteration.

Depending on the quality of images that are projected through the scattering medium (e.g. a fiber), this process could be re- peated for multiple times. It is expected that after each itera- tion, the domain of the SLM patterns generated by the second set of networks G gets closer to the domain of SLM images that pro duce perfect target images at the output of the scattering medi um (e.g. a fiber) . It should be noted that the similarity of the proposed networks to GAN' s architecture roots from the fact that training of the second set of networks G, i.e. mapping from am plitude patterns to SLM images, cannot be straightforwardly car ried out because no label (ground truth SLM images) for target output images exists a-priori (training cannot be performed in a supervised manner in which ground truth labels are available be forehand) . Therefore, the performance of the second set of net works G gets better by competing with the first set of networks D where it tries to generate SLM images that result in output amplitude images with higher fidelities; hence the name genera tive adversarial network.

In another embodiment of the present invention, the first set of networks proposed in the previous embodiment that emulates the forward path of light propagation through said scattering medium can itself be comprised of two sub-sets, one sub-set emulating the real and the other sub-set emulating the imaginary part of the complex input patterns coming from the light modulator and propagating to the screen. Similarly, the second set of networks that emulates the reverse path of light propagation through said scattering medium, can also be comprised of two sub-sets, one for emulating the real and one for emulating the imaginary part of light propagation from the screen to the light modulator. The training of the sets of networks in the forward and reverse path, which includes the training of the sub-networks, is exact- ly carried out in the same fashion as described in the previous embodiment .

In another embodiment of the present invention, a colored RGB image can be projected through the scattering medium by using three different light sources for each color (corresponding to a specific wavelength) and used for an approach similar to the previous embodiments. For example, but not limited to, in the case of the neural network method similar to the GAN architec ture, the architecture of the network could be changed in a way that it consists of three different training units for each col or. The procedure for projecting an image with each color is then carried out in the same fashion as in the previous embodi ments. Once trained, each unit generates its own SLM pattern. Then, for display on the screen, a rapid cycle through the three SLM patterns transmitted through the scattering medium and ob served on the screen gives the impression of a colored image to the human eye. The three light sources could be for example, but not limited to, three different laser sources with the red, green and blue colors, or one tunable laser, or three or more LEDs with appropriate emitting wavelengths. The incoherency of the latter source, i.e. LEDs, is also beneficial for suppressing the unwanted background of the projected images due to shortcom ings of the neural network in generating the appropriate SLM pattern or incomplete light modulation by the SLM (amplitude- only or phase-only modulations instead of amplitude-and-phase (complex) modulation) .

BRIEF DESCRIPTION OF THE DRAWINGS These and other features, aspects and advantages of the present invention will become better understood with regard to the fol lowing non-limiting examples and accompanying drawings where:

Figure 1 is a schematic illustration of an image projector device that modulates the incoming light with the proper modulation to project images of certain shapes on the screen.

Figure 2 : is a schematic illustration of an embodiment of obtaining input-output data to train the neural networks in both a forward way and reverse way.

Figure 3: is a schematic illustration of the neural networks and its training procedure in which the network learns to generate proper modulation that when up loaded on the light modulator, projects certain pattern on the screen. Figure 4 : shows examples of the projected images on the

screen using amplitude-only modulation for input SLM patterns. The examples are Latin alphabet characters .

Figure 5: shows examples of projected images on the screen using amplitude-only modulation for input SLM pat terns. The examples are digit images,

Figure 6: shows examples of projected images on the screen using amplitude-only modulation for input SLM pat terns. The examples are drawings of a running per son .

Figure 7 : shows examples of projected images on the screen using amplitude-only modulation for input SLM pat terns. The examples are drawings of a skull- shape, a heart, a smiley and a circle,

Figure 8 : shows examples of projected images on the screen for three different wavelengths (488nm, 532nm, 633nm) , RGB colored images and gray-scale images of incoherent superposition (sum) of three wave lengths projected on the screen using amplitude and phase (complex) modulation for input SLM pat terns. The examples are Latin alphabet characters,

Figure 9: shows examples of the projected RGB images on the screen as well as incoherent superposition (sum) of three different wavelengths (488nm, 532nm, 633nm) projected on the screen using amplitude and phase (complex) modulation for input SLM patterns generated using the proposed neural network ap proach as well as the full measurement transmis- sion matrix approach (provided for the sake of comparison) . The examples are comprised of vari ous drawings .

Figure 10: shows examples of the projected images on the

screen for three different wavelengths (488nm, 532nm, 633nm) , RGB colored images and gray-scale images of incoherent superposition (sum) of three wavelengths projected on the screen using ampli tude and phase (complex) modulation for input SLM patterns. The examples are portraits of people and cartoon characters.

EXEMPLARY EMBODIMENTS

In figure 1, a schematic illustration of an embodiment of a mul- timode optical fiber projector 1 according to the present inven tion is shown. The projector comprises a laser source 2, such as a superluminescent diode, SLED or a Light emitting diode (LED) , an optional lens 2a provided between the laser source 2 and a subsequent light modulator 3, a light modulator 3 selected from the group consisting of a spatial light modulator (SLM) and a digital micromirror device (DMD) ) , a device 4 comprising a scat tering medium, such as a multimode fiber (MMF) , and a camera system comprising a screen (5) . Wavefront shaping is required to undo the scrambling of the scattering medium (e.g. the MMF) . This wavefront shaping is carried out by a system of two or more neural networks (not shown in Fig. 1) . The system of networks is given the desired image that is to be projected on the screen 5, and the appropriate wavefront (i.e. spatial modulation pattern) is computed by the system of networks.

In figure 2, a schematic illustration of an embodiment of the optical setup for acquiring examples for forward and reverse training of the system of networks is presented. For training the system of networks in the forward path of light transmission inside the scattering medium 4, a set of many examples from the original SLM 3 to the original screen 5 is provided. For this, light from a laser source 2 is led through a beam splitter 6 to the original SLM 3, where it is modulated and sent through the light scattering medium 4 (here an MMF) to another beam splitter 6', from where it is led via mirrors 7, Ί' to the original screen 5. The system of networks is then trained with this set of examples. For training the networks that are intended to emu late the reverse path of light propagation, examples relating to the pattern on the secondary SLM 3' and the secondary screen 5' are obtained and the networks are trained with the set of these examples. For this, the light from the laser source 2 is led through the beam splitter 6, from where it is sent through the light scattering medium 4 (here an MMF) to another beam splitter 6', from where it is led via mirror 1 to a secondary (extra) SLM 3' . There, the light is modulated and sent back in the re verse path direction via mirror 1’’ , beam splitter 6', through the scattering medium (MMF) 4 to the beam splitter 6, from where it is led via mirror 7 ^{, , ,} to a secondary (extra) screen 5'. In figure 3, the overall schematic of the projector network sys tem comprising two sub-networks is shown: a discriminator (D) (first set of networks) and a generator (G) (second set of net works) are shown. Once trained, the sub-network G accepts the target pattern desired to be projected at the output of the MMF (i.e. the desired image) and accordingly generates an SLM image corresponding to that pattern. The role of the sub-network D, which is trained to emulate the optical forward path SLM-MMF- camera, is to help the network G in the training step to come up with SLM patterns which result in projected images on the camera with higher fidelities. In this diagram, the fiber, the discrim inator and generator sub-networks are denoted by F, D and G, re spectively. The training procedure is carried out in three steps :

A- a number of experimental examples from the SLM images to the amplitude-only images on the screen are obtained. B- the sub-network D is trained on these images to learn to map the SLM images to screen images; hence D is essentially learning the optical forward path SLM-MMF-camera .

C- the network weights of D are fixed and sub-network G is trained with target images as input and produces an SLM image corresponding to the target image. The SLM image is then passed to the fixed sub-network D which produces an estimate of the fi ber output (NN output) based on the SLM image it received. The error between the output of D and the target image (desired im- age) (input of G) is back propagated to G to update its traina ble weights and biases. The validation procedure is carried out by feeding the target image (desired image) to the trained sub network G and acquiring the appropriate SLM image corresponding to that target image.

Figure 4 shows examples of the projected images from the Latin alphabet; (a) train and (b) test dataset captured on the camera are shown. The target image, shown on the bottom right corner of each image, is fed to the neural network to generate the appro- priate SLM pattern (amplitude-modulated) that produces that de sired target image corresponding to the central part of the fi ber's facet FOV (field of view) . This area is shown as a dashed box on the top left image. The fidelity of the projected image with respect to its corresponding grayscale target image is shown as inset in each image. In figure 5, examples of the projected images from the MNIST handwritten digits dataset are provided. Target images, shown next to the projected images, are blindly fed to the neural net- work system which was trained only on alphabet characters. The neural network system-generated SLM patterns (amplitude- modulated) are sent through the fiber and the outputs of the fi ber are captured on the screen and by a camera. The fidelity of the projected image with respect to its corresponding grayscale target image is shown as inset in each image.

Figure 6 shows the result of the neural network system trained on Latin alphabet characters when used to generate the required SLM pattern (amplitude-modulated) that produces a drawing of a running person in an area of 200x200 pixels (1.6x1.6 mm2) corre sponding to the central part of the fiber's facet FOV. The de sired target image (left), the corresponding SLM image (middle) and the projected image (right) are depicted. The fidelity of the projected image with respect to its desired image is shown as inset.

In figure 7, examples of the projected random drawings (a skull a heart, a smiley, a circle) are provided. Target images, shown in bottom right corner, are blindly fed to the neural network system which was trained only on alphabet characters. The neural network system-generated SLM patterns (amplitude-modulated) are sent through the fiber and the outputs of the fiber are captured on the camera. The fidelity of the projected image with respect to its corresponding grayscale target image is shown as inset in each image.

In figure 8, examples of the projected colored images from the EMNIST Latin characters dataset for three different wavelengths (488nm, 532nm, 633nm) are provided. Target images are separately fed to the neural network system which was trained only on al phabet characters for a single color. The neural network system generated SLM patterns (amplitude and phase modulated) are sent through the fiber and the outputs of the fiber corresponding to the central part of the fiber's facet FOV (shown as a dashed box on in one of the images) are captured on the screen and by the camera. The three colored images can be recombined as a stack of images comprised of three channels (one channel for each color) to create a single colored RGB image. Similarly, the three imag es can be superimposed to create a single incoherent gray-scale image. The fidelity of the projected images with respect to their corresponding grayscale target images is shown as inset in each image.

In figure 9, various examples of the projected colored RGB imag- es comprising of three wavelengths (488nm, 532nm, 633nm) as well as the incoherent superposition thereof are provided. Target im ages are fed to the system of neural networks trained only on alphabet characters. The neural network system-generated SLM patterns (amplitude and phase modulated patterns) are sent through the fiber and the outputs of the fiber are captured on the screen and by the camera. For the sake of comparison, simi lar images are projected through the fiber using the convention al transmission matrix approach. In figure 10, various examples of the projected colored RGB im ages comprising of three wavelengths (488nm, 532nm, 633nm) as well as the incoherent superposition thereof are provided. Tar get images comprising of pictures of human portraits are fed to the system of neural networks trained on these images. The neu- ral network system-generated SLM patterns (amplitude and phase modulated) are sent through the fiber and the outputs of the fi ber are captured on the screen and by the camera.