A method and system for dynamic holographic generation of a reconfigurable diffractive neural network

Abstract

The application discloses a kind of dynamic holographic generation methods and systems of reconfigurable diffractive neural network, it is related to optical computing and optical information processing technical field, comprising: S1, input image is loaded to input plane using amplitude coding;S2, construct including multiple passive diffraction layers and multiple reconfigurable diffractive layers diffractive optical network;S3, main light source is irradiated input plane by uniform plane wave, generates input light field sent into passive diffraction layer;S4, liquid crystal spatial light modulator LC-SLM obtains key light under the irradiation of auxiliary light source;S5, by key light, write diffractive modulation phase into each reconfigurable diffractive layer and modulate, decryption operation, to generate phase hologram in real time dynamically in physical diffraction propagation process;S6, optical detector completes the optical reconstruction of target light field on detection surface.The application introduces light modulation reconfigurable diffractive layer for the first time, allows to adjust phase distribution in real time by external key light signal, realizes all-optical real-time holographic encoding / decoding.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of optical computing and optical information processing. More particularly, the present application relates to a dynamic holographic generation method and system of a reconfigurable diffractive neural network. BACKGROUND

[0002] Since its inception in the 1960s by Kozma and Kelly, Computer Generated Holography (CGH) has become an important pillar in the field of optical imaging and display. This technology directly generates the object light information of a virtual object through numerical calculation, without the need for traditional holographic recording processes, thereby achieving precise control of the light field. In principle, CGH uses Fourier transform or other diffraction integral algorithms (such as Fresnel or Fraunhofer diffraction models) to simulate the propagation and interference of object light waves, forming a hologram, and then reproducing a three-dimensional image through devices such as spatial light modulators (SLMs). This method has shown great potential for application in fields such as three-dimensional display, data storage, optical encryption, and biomedical imaging. For example, in near-eye display systems, CGH can generate stereoscopic images with natural depth of field, improving visual comfort; while in lithography and wavefront modulation applications, it can achieve light field manipulation under complex media. However, as application scenarios become more complex, CGH faces increasingly severe challenges, particularly in terms of real-time requirements.

[0003] Existing CGH techniques mainly rely on iterative algorithms (such as the Gerchberg-Saxton algorithm, the Fienup phase retrieval algorithm, or non-convex optimization methods) and deep learning models (such as U-Net, ResNet, or generative adversarial networks, GAN). These methods can generate high-quality phase-only holograms (POH), but the calculation process is time-consuming, relies on high-performance electronic devices, and faces energy consumption and delay problems in real-time applications. For example, the Gerchberg-Saxton algorithm iteratively optimizes the phase distribution through multiple Fourier transforms, but for high-resolution (such as 4K or higher) hologram generation, it often takes several seconds or even minutes to calculate, which cannot meet the needs of dynamic scenarios. In deep learning methods, although the research team at MIT has achieved accelerated hologram generation by using neural networks to learn physical laws through "tensor holography", it still requires electronic processors such as GPUs for inference, which introduces significant power consumption (for example, the training and inference process may consume several watts of power). In addition, when dealing with massive point cloud data or real-time moving images, traditional CGH is difficult to expand, leading to limited application in augmented reality (AR) and virtual reality (VR), and users may experience visual fatigue or dizziness caused by delay. In recent years, with the popularity of AR / VR devices (such as head-mounted displays), the demand for real-time performance has become increasingly urgent: in secure communication, it is necessary to generate encrypted holograms instantaneously to prevent information leakage; in medical imaging, it is required to reconstruct three-dimensional biological tissue images in real time to support surgical navigation; in quantum communication, holographic technology needs to be combined with entangled photons to achieve high-bandwidth, low-latency data transmission. These needs highlight the bottlenecks of existing electronic-driven CGH: high computational complexity (O(N^2 log N) level, N is the number of pixels), high energy consumption (electronic chip heat dissipation problem), and poor adaptability to dynamic input.

[0004] In recent years, all-optical diffractive neural networks (DNNs) have emerged as a new optical computing framework, achieving light-speed holographic generation through passive diffraction layers, thus avoiding the bottleneck of digital computing. This framework simulates the hierarchical structure of neural networks, utilizing diffraction propagation (such as the angular spectrum method ASM) for parallel computation in free space, achieving POH generation speeds at the picosecond level, with energy consumption limited to the illumination source (typically <1mW). For example, the DNN system developed by VividQ based on a hierarchical FFT algorithm has demonstrated its potential in AR displays, capable of projecting highly realistic 3D scenes. However, existing DNN frameworks have fixed parameters, and once trained, they cannot adapt to new input types or dynamic scenes, resulting in insufficient flexibility in practical deployments. Furthermore, in security-sensitive applications, such as data encryption or privacy protection, existing holographic systems lack effective encryption mechanisms and are vulnerable to reverse engineering attacks. For example, POH generated by static DNNs can be reverse-engineered, exposing the original object information; wavefront modulation in complex media (such as scattering structures) also faces noise interference and security challenges. Summary of the Invention

[0005] One object of the present invention is to solve at least the above-mentioned problems and / or defects, and to provide at least the advantages described below.

[0006] To achieve these objectives and other advantages of the present invention, a method for dynamic holographic generation of a reconfigurable diffraction neural network is provided, comprising:

[0007] S1. Load the input image onto the input plane after using amplitude encoding;

[0008] S2. Construct a diffraction optical network comprising multiple passive diffraction layers and multiple reconfigurable diffraction layers;

[0009] S3. The main light source illuminates the input plane with a uniform plane wave, generating an input light field that is fed into the passive diffraction layer. Each passive diffraction layer diffracts and modulates the input light field layer by layer.

[0010] S4. After the computer terminal loads the encrypted phase mask, it outputs it to the liquid crystal spatial light modulator (LC-SLM), and obtains the key light under the illumination of the auxiliary light source.

[0011] S5. The key light is used to write the diffraction modulation phase with characteristic distribution into each reconfigurable diffraction layer, and the target light field is further modulated and decrypted to generate a phase hologram in real time during the physical diffraction propagation process.

[0012] S6. The optical detector captures the intensity distribution corresponding to the phase hologram on the detection surface, thus completing the optical reconstruction of the target light field;

[0013] Wherein, for different holographic computing tasks, the reconfigurable diffraction layer can change the phase parameters of the diffraction modulation layer through the phase mask irradiated by the auxiliary light source, so as to realize the dynamic adjustment of the holographic task.

[0014] Preferably, in S1, the initial size of the input image is 28*28 pixels, which is expanded to 32*32 pixels by bicubic interpolation, then spliced into a custom image of 96*96 pixels, and the custom image is expanded to 500*500 pixels by zero padding to match the input field of view.

[0015] The input image matching the input field of view is adjusted to an amplitude image and printed on a glass substrate to obtain an input plane.

[0016] Preferably, in S2, the passive diffraction layer is constructed in the following manner:

[0017] S20, processing the picture of the data set into a training set and a test set matching the input field of view;

[0018] S21, defining the diffractive optical network as a phase randomly initialized in the range of [0, 2π], adding a dynamic interface on the reconfigurable diffraction layer, limiting the phase with a sigmoid function, multiplying the input image with the plane wave amplitude to define the forward propagation, simulating the diffraction propagation by the angular spectrum method, and completing the parameter initialization of the diffraction neural network;

[0019] S22, using the pictures in the training set as input, generating encrypted POH through forward propagation, calculating the loss after decoding and reconstructing the image, updating the parameters after back propagation to complete one round of training;

[0020] S23, after a predetermined number of training rounds, the solidified parameters of the passive diffraction layer are obtained.

[0021] Preferably, in S3, the diameter of the uniform plane wave is 10 cm, and the uniformity is greater than 95%.

[0022] A dynamic holographic generation system applied to the dynamic holographic generation method of the reconfigurable diffraction neural network, comprising:

[0023] An input plane containing amplitude encoding of an input image;

[0024] A main light source cooperating with the input plane;

[0025] A diffractive optical network arranged downstream of the input plane, the diffractive optical network comprising: a plurality of passive diffraction layers with fixed phase distribution, a plurality of reconfigurable diffraction layers arranged at the end of the passive diffraction layers and composed of optical phase change material;

[0026] An auxiliary light source cooperating with each reconfigurable diffraction layer;

[0027] a liquid crystal spatial light modulator (LC-SLM) in communication with the computer terminal;

[0028] an optical detector arranged downstream of the diffractive optical network for receiving the modulated light field by the diffractive optical network and recording the output hologram.

[0029] Preferably, the main light source adopts He-Ne laser I with a wavelength of 532 nm as the light source, and the power of the He-Ne laser I is 5 mW.

[0030] Preferably, the number of layers of the passive diffractive layer is 3 layers; and the number of layers of the reconfigurable diffractive layer is 2 layers.

[0031] Preferably, the distance between the input plane and the first layer of the passive diffractive layer is 12 cm, the distance between adjacent layers of the diffractive optical network is 20 cm, and the distance between the terminal reconfigurable diffractive layer and the optical detector is 15 cm.

[0032] Preferably, the diffractive optical network is arranged between the input plane and the optical detector by a support, and the axial alignment error between the diffractive optical networks is less than 0.1 mm.

[0033] Preferably, the auxiliary light source adopts He-Ne laser II with a wavelength of 633 nm and a power of 1 mW, and the output light of the He-Ne laser II is output to the reconfigurable diffractive layer through a fiber coupler and a polarization controller.

[0034] The present application at least has the following beneficial effects: the core idea of the present application is to expand the static diffractive neural network into a dynamic reconfigurable framework, and for the first time, the phase control of the reconfigurable layer is realized by introducing the method of regulating the optical phase change material by the auxiliary light source (i.e. the cooperation of the auxiliary light source + LC-SLM + reconfigurable layer), which allows the phase distribution of the reconfigurable diffractive layer to be adjusted in real time by an external key light signal, and realizes real-time holographic encoding / decoding in an all-optical manner.

[0035] The present application is different from the traditional DNN with fixed parameters, and the present system supports online fine-tuning, adapts to dynamic input (such as real-time video or variable environment objects), and the calculation speed reaches the level of light speed. The passive layer does not need external power supply, and only the reconfigurable layer needs a small amount of light control, and the energy consumption is much lower than that of the electronic neural network.

[0036] Other advantages, objects, and features of the present application will be apparent from the following description, and will be understood by those skilled in the art. BRIEF DESCRIPTION OF DRAWINGS

[0037] Figure 1 a schematic diagram of the dynamic holographic generation system of the present application;

[0038] Wherein, the main light source-1, input plane-2, diffractive optical network-3, passive diffractive layer-30, reconfigurable diffractive layer-31, auxiliary light source-4, optical detector-5, liquid crystal spatial light modulator-6. DETAILED DESCRIPTION

[0039] The application will be further described in detail below with reference to the accompanying drawings, so that those skilled in the art can implement the application according to the description and the drawings.

[0040] A dynamic holographic generation system based on a reconfigurable diffractive neural network, by introducing a reconfigurable diffractive layer (light control phase change material), combined with external light signals, realizes real-time reconfiguration of the network. The system not only retains the parallelism and low energy consumption advantages of all-optical computing, but also has dynamic adaptability, making the system suitable for real-time holographic display, quantum communication and biomedical imaging fields. Compared with traditional methods, the application significantly improves the robustness and task multiplexing ability of the system, filling the gap in the dynamic encoding field of all-optical holography, such as Figure 1 As shown in the figure, the dynamic holographic generation system comprises:

[0041] A main light source 1 for generating a uniform plane wave to irradiate an input plane to generate an input light field sent to a passive diffractive layer, the main light source uses a He-Ne laser I with a wavelength of 532nm as a light source, and the power of the He-Ne laser I is 5mW;

[0042] Input plane 2: Place the image of the object to be encoded, irradiated by a uniform plane wave.

[0043] Diffractive optical network 3: composed of multiple passive diffractive layers 30 and at least one reconfigurable diffractive layer 31, each layer contains tens of thousands of diffractive features (neurons) trained by self-supervised learning combined with reinforcement learning. The system optimizes network parameters through hybrid training, supports extended depth focus (6-15cm). When encrypted, the network generates noise POH.

[0044] Auxiliary light source 4 for writing the phase distribution obtained by training into the phase change optical material to form a reconfigurable diffractive modulation layer with characteristic distribution for further modulation of the target light field, the auxiliary light source uses a He-Ne laser II with a wavelength of 633nm and a power of 1mW, the refractive index change of the modulation layer is triggered by the auxiliary light source, and the phase of the diffractive layer is dynamically adjusted in real time, so as to adjust the phase distribution of the reconfigurable diffractive layer;

[0045] Liquid crystal spatial light modulator 6 (liquid crystal spatial light modulator is abbreviated as LC-SLM) in communication connection with the computer terminal, for writing the predetermined decryption information into the light wave under the cooperation of the auxiliary light source, and modulating the light wave to generate a key light, when the key light is transmitted to the reconfigurable diffractive layer 31, it can cooperate with the light splitting mechanism and the reflection mechanism to realize the light splitting transmission of the reconfigurable diffractive layer 31.

[0046] The optical detector 5 is arranged at the location for receiving the light field intensity distribution modulated by the diffraction neural network;

[0047] The optical signal output by the optical detector 5 is input to the computing unit through the data acquisition and processing module, and the computer algorithm is used to decode or encode the light field to generate specific phase optical holographic (POH) data.

[0048] Working principle: the output light of He-Ne laser II illuminates the DMD spatial light modulator loaded with specific patterns to generate optical patterns, which adjust the phase distribution of the reconfigurable diffraction modulation layer in the form of intensity distribution. The optical information on the input plane is diffracted and forward propagated under the illumination of the main light source, input to the first diffraction modulation layer, and then all neurons on the first diffraction modulation layer transmit optical signals to the next diffraction modulation layer in the form of light field diffraction propagation, and finally the output light field is transmitted to the reconfigurable diffraction layer (which is configured to be irradiated by the light reflected or transmitted by the liquid crystal spatial light modulator LC-SLM, and the optical phase change material can dynamically adjust the phase distribution according to the optical signal), and the dynamic modulation is realized through controlled propagation. In this way, free space diffraction transmission and controlled modulation can be realized.

[0049] In specific implementation, the present application also relates to a specific implementation method, covering system construction, training, encoding and decoding process, specifically including the following processing stages:

[0050] I. System hardware construction stage

[0051] Step 1.1: Prepare the input plane 2, use the amplitude encoded object image (initial size 28x28 pixels, expand to 32x32 by bicubic interpolation, then splice to 96x96 custom image, and finally zero pad to 500x500 to match the input field of view). The image is illuminated by a uniform plane wave with a wavelength of 532 nm (using He-Ne laser as the main light source 1, power about 5 mW, ensuring uniformity > 95%).

[0052] Step 1.2: Constructing the diffractive optical network 3, including 5 layers (the first 3 layers are passive diffractive layers 30, made by 3D printing or lithography technology using a fixed phase mask, with a phase range [0, 2π], passive diffractive layers 30 use a fixed phase that cannot be dynamically changed, which can be obtained using liquid crystals or fixed metasurfaces, such as using a liquid crystal spatial light modulator LC-SLM, such as the Hamamatsu X10468 series, with a resolution of 600x600 pixels and a pixel size of 8μm; the last 2 layers are reconfigurable diffractive layers 31, which use optical phase change materials and use auxiliary optical modulation to adjust the phase). Each layer in the diffractive optical network contains 600x600 neurons, each neuron has a size of 8μm. In this step, the diffractive optical network composed of passive diffractive layers 30 and reconfigurable diffractive layers 31 is used to generate a phase hologram (POH) in real time, and then the reconfigurable layer is dynamically adjusted by an external light signal to realize the full-optical holographic encoding of the input target and the reconstruction of the target image, with the advantages of extended depth focus, light speed calculation, and high security.

[0053] Step 1.3: Set the interlayer distance to 12-20cm (specifically, 12cm to the first layer, 20cm between adjacent layers, and 15cm from the last layer to the output plane). Use a precision optical support (such as a Thorlabs optical platform) to fix the layer position and ensure that the axial alignment error is <0.1mm.

[0054] Step 1.4: Integrate the key light control module, use auxiliary light source 4 (wavelength 633nm, power 1mW, polarization control achieved by half-wave plate) to irradiate the reconfigurable layer. The control interface is connected to a computer or FPGA board to realize pulse sequence adjustment (phase range [0, 2π], response time <10ms).

[0055] Step 1.5: Prepare the decoding module, use the Fresnel diffraction propagation model (free space propagation distance 15cm, variable range 6-15cm). The optical detector 5 on the output plane uses a CCD camera (such as Thorlabs DCC1545M, resolution 1280x1024) to capture the POH.

[0056] II. Training phase of diffractive optical network theory

[0057] This phase uses numerical simulation for training to achieve diffractive network parameter optimization, detailed to load data sets, ensure repeatability, in specific implementation, mainly use self-supervised learning combined with reinforcement learning for training, support dynamic input adaptation, such as video frame sequence or variable environment object, specific processing steps include:

[0058] Step 2.1: Environment preparation. Use Python 3.9+ and PyTorch 1.9 framework, install necessary libraries (numpy, torch). Set up GPU acceleration (e.g. NVIDIA RTX 30 series) with random seed 42 for reproducibility.

[0059] Step 2.2: Load dataset. Download Fashion-MNIST dataset. Preprocessing: bicubic interpolation of 28x28 images to 32x32 (using PIL.Image.resize); randomly tile 9 32x32 images into a 96x96 custom image; further interpolation to 500x500 (cv2.resize); zero-padding to 600x600 to match input field of view (FOV). Create training set (12,000 fashion product images) and test set (2,000 fashion + 2,000 handwritten).

[0060] Step 2.3: Initialize network parameters. Define diffraction layer: passive layer phase randomly initialized [0, 2π]; reconfigurable layer similar but with added dynamic interface. Use sigmoid function to limit phase. Define forward propagation: input image multiplied by plane wave amplitude; simulate diffraction propagation via angular spectrum method (ASM) (torch.fft.fft2 computes frequency spectrum, multiplied by transmission function exp(j2π / lambda* sqrt(1 - (fx^2 + fy^2)*lambda^2)), where lamda is wavelength, j represents imaginary unit, fx, fy represent spatial frequency, torch.fft.ifft2 inverse transform).

[0061] Step 2.4: Define loss function. Hybrid loss: NPCC (negative Pearson correlation coefficient) + MSE (mean squared error) + encryption strength loss (mutual information, estimated using torch mutual information). Introduce random axial shift Δz ~ U(-5cm, 5cm) and key perturbation (random phase α*R, α=0.1, R~[0, 2π]).

[0062] Step 2.5: Training cycle. Input the number of training data subsets batch (e.g., 20 images) used at each parameter update; forward propagation to generate POH; decoding to reconstruct images; loss calculation; backpropagation (using Adam optimizer, set learning rate lr = 0.001, learning rate decay decay = 0.99epoch * 1e-3, epoch is a complete cycle in neural network training, that is, the process of inputting the entire training data set into the model once and completing forward and backward propagation); update parameters. Monitor indicators: calculate test set SSIM / PSNR every 10 epochs. After training, solidify the passive layer parameters (save as a.npy file for physical production), and keep the reconfigurable layer dynamic optimization.

[0063] Step 2.6: Evaluation and generalization. Calculate the average SSIM > 0.4, PSNR > 10 dB, and mutual information < 0.1 using the test set. Test external generalization (handwritten digits).

[0064] III. Phase pattern generation stage

[0065] After the training of the diffractive optical network theory, it is applied to step 1.2 to generate the encoded POH in real time. Compared with existing static optical networks or digital holographic methods, this stage provides real-time adaptive coding, improving the application potential of optical computing in secure communication, augmented reality (AR) privacy protection, and medical image encryption. The specific steps are as follows:

[0066] Step 3.1: Input unknown object images at the input plane, generate input light fields sent to the passive diffractive layer after illumination by the plane wave of the main light source, and propagate and modulate through the passive layer (light field diffracts layer by layer).

[0067] Step 3.2: The reconfigurable layer forms a characteristic distribution diffractive modulation layer under the illumination of the auxiliary light source to further modulate the target light field.

[0068] Step 3.3: CCD captures the output plane intensity distribution at the detection plane, and converts it into POH (sigmoid(I)*2π) according to the quantitative phase imaging method, where I is the light intensity distribution of the detection plane. This phase generation process is completed at full optical speed.

[0069] IV. Optical reconstruction stage

[0070] This stage mainly uses physical diffraction propagation to reconstruct images, and the specific processing steps include:

[0071] Step 4.1: Load the encrypted phase mask to the output liquid crystal spatial light modulator LC-SLM on the computer terminal, use coherent light to illuminate the liquid crystal spatial light modulator LC-SLM to make the encoded phase information forward transmission, and get the key light.

[0072] Step 4.2: Reconfigurable layer phase adjustment by key light (response time <10 ms), light field propagation through adjustment network, generating decrypted POH.

[0073] Step 4.3: Reconstructing in image plane using Fresnel propagation (CCD captures intensity, distance 6-15 cm adjustable). Evaluation: SSIM / PSNR.

[0074] Five, optical experiment deployment steps

[0075] This stage describes the prototype construction and testing in detail, with very detailed deployment details to ensure safety and precision, and specific processing steps include:

[0076] Step 5.1: Laboratory environment preparation. Use an optical darkroom (temperature 20-25°C, humidity <50%, dust level ISO7), install a vibration isolation optical table (Thorlabs PTS603, size 1m x 1m). Power supply: stable laser power supply (<1% fluctuation). Safety measures: wear laser safety glasses (OD4+ for 532 / 633nm), set up laser barriers.

[0077] Step 5.2: Light source deployment. Main light source: 532nm He-Ne laser (Thorlabs HNL210L, power 5mW), generate uniform plane wave (diameter 10cm, uniformity >95%) through beam expander (f=50mm) and collimating lens (f=200mm). Key light source: 633nm He-Ne laser (power 1mW), coupled through fiber coupler and polarization controller (e.g. use manual fiber polarization controller Thorlabs FPC 560, adjust polarization through half-wave plate).

[0078] Step 5.3: Input plane deployment. Use a transparent mask (glass substrate printed with amplitude image, resolution >1000dpi) placed on the optical table, 10cm from the laser output. Image alignment: use laser alignment instrument to ensure center axis alignment error <0.05mm.

[0079] Step 5.4: Diffractive network deployment. Passive layer: use 3D printer (Formlabs Form 3, material ClearResin) to make phase mask (convert STL model from training.npy file, thickness 0.5mm, phase accuracy <λ / 10). Reconfigurable layer: use optical phase change material (based on GST / GSST film), write phase programmable control through auxiliary light source.

[0080] Layer fixing: Precise stand (Thorlabs KM100) used, distance measured with laser rangefinder (precision 0.01 mm). Axial alignment: He-Ne laser point used layer by layer, far-field diffraction pattern observed to adjust to symmetry.

[0081] Step 5.5: Output and decoding deployment. Output plane: CCD camera (Thorlabs DCC1545M, exposure time 1 ms, gain 0 dB) placed, 15 cm from last layer. Decoding: free-space propagation, distance changed using adjustable stand (6-15 cm in steps of 1 cm). Captured images transferred to PC for processing (Matlab or Python for SSIM / PSNR analysis).

[0082] Step 5.6: Control system deployment. PC (Intel i7, 16 GB RAM) running control software: Python script loads key sequence, SLM driver API adjusts phase. Synchronization: Arduino board used to trigger laser pulses and CCD capture (delay <1 ms).

[0083] Step 5.7: Testing and calibration.

[0084] Initial calibration: background noise measured with no input (<1% of signal). Encryption test: output POH captured with no key applied to input image, and mutual information calculated between original image and captured hologram to verify system secrecy.

[0085] Decryption test: optical key applied to adjust reconfigurable layer phase distribution, image reconstructed, and distance test depth focus (SSIM curve recorded).

[0086] Data recording and output: system automatically saves original input image, output hologram, decrypted image, and related performance indicators (including mutual information, SSIM curve, attack failure rate, etc.) for each test, and generates a timestamped test log file for traceability.

[0087] Step 5.8: Troubleshooting and optimization.

[0088] Common problems: alignment offset (re-calibration); SLM response slow (reduce power); high noise (increase filter mirror).

[0089] Optimization: key sequence iteratively adjusted to maximize encryption strength, where key sequence represents an ordered set of phase / amplitude encoding matrices loaded onto reconfigurable diffractive layers (e.g. SLM / DM / DMD) in "layer sequence and time sequence" during encryption / decryption process, along with key stream composed of its loading order and timing parameters. Specifically, for an L-layer reconfigurable diffractive network operating at T time steps, key sequence S is defined as:

[0090]

[0091] wherein, is the complex amplitude transmission function (phase mask) loaded at the lth layer at the tth time step, is the loading time for this step.

[0092] The above-described solution is only a preferred example, but is not limited thereto. In implementing the present application, appropriate substitutions and / or modifications can be made according to the user's needs.

[0093] While the embodiments of the present application have been disclosed as above, it is not limited to the uses listed in the specification and the embodiments. It can be fully applied to various fields suitable for the present application. Further modifications can be easily made by those skilled in the art. Therefore, the present application is not limited to the specific details and the figures shown and described herein without departing from the general concept defined by the claims and the equivalent scope.

Claims

1. A method for dynamic holographic generation of a reconfigurable diffractive neural network, characterized in that, Comprising: S1, loading the input image to the input plane after amplitude coding; S2, constructing a diffractive optical network comprising a plurality of passive diffraction layers and a plurality of reconfigurable diffraction layers; S3, the main light source irradiates the input plane with uniform plane waves to generate input light field into the passive diffraction layer, and each passive diffraction layer carries out layer-by-layer diffraction propagation and modulation on the input light field; S4, the computer terminal loads the encrypted phase mask and outputs to the liquid crystal spatial light modulator LC-SLM, and obtains the key light under the irradiation of the auxiliary light source; S5, write the diffraction modulation phase with characteristic distribution into each reconfigurable diffraction layer through the key light, and further modulate and decrypt the target light field to generate phase hologram in real time during the physical diffraction propagation; S6, the optical detector captures the intensity distribution corresponding to the phase hologram on the detection surface to complete the optical reconstruction of the target light field; Wherein, for different holographic computing tasks, the reconfigurable diffraction layer can change the phase parameters of the diffraction modulation layer through the phase mask irradiated by the auxiliary light source, so as to realize the dynamic adjustment of the holographic task.

2. The method of claim 1, wherein, In S1, the initial size of the input image is 28×28 pixels, which is expanded to 32×32 pixels by bicubic interpolation, then spliced into a custom image of 96×96 pixels, and expanded to 500×500 pixels by zero padding to match the input field of view; After adjusting the input image matching the input field of view to an amplitude image, it is printed on a glass substrate to obtain an input plane.

3. The method of claim 1, wherein, In S2, the construction method of the passive diffraction layer comprises: S20, process the data set pictures into training set and test set matching the input field of view; S21, define the diffractive optical network as phase random initialization in the range of [0, 2π], add a dynamic interface on the reconfigurable diffraction layer, limit the phase with sigmoid function, define the input image multiplied by plane wave amplitude as forward propagation, simulate the diffraction propagation by angular spectrum method, and complete the parameter initialization of the diffractive neural network; S22, use the pictures in the training set as input, generate encrypted POH through forward propagation, calculate the loss after decoding and reconstructing the image, update the parameters after back propagation to complete one round of training; S23, after a predetermined number of training, the solidified parameters of the passive diffraction layer are obtained.

4. The method of claim 1, wherein, In S3, the diameter of the uniform plane wave is 10 cm, and the uniformity is greater than 95%.

5. A dynamic holographic generation system applied to the dynamic holographic generation method of the reconfigurable diffractive neural network according to any one of claims 1-4, characterized in that, Comprising: An input plane containing amplitude coding of an input image; A main light source matched with the input plane; A diffractive optical network arranged downstream of the input plane, the diffractive optical network comprising: a plurality of passive diffraction layers with fixed phase distribution, a plurality of reconfigurable diffraction layers arranged at the end of the passive diffraction layers and composed of optical phase change material; An auxiliary light source matched with each reconfigurable diffraction layer; A liquid crystal spatial light modulator LC-SLM in communication connection with a computer terminal; An optical detector arranged downstream of the diffractive optical network for receiving the light field modulated by the diffractive optical network and recording the output hologram.

6. The dynamic holographic generation system of claim 5, wherein, The main light source uses He-Ne laser I with a wavelength of 532 nm as the light source, and the power of He-Ne laser I is 5 mW.

7. The dynamic holographic generation system of claim 5, wherein, The passive diffraction layer has 3 layers; and the reconfigurable diffraction layer has 2 layers. The distance between the input plane and the first layer of passive diffraction layer is 12 cm, the distance between the adjacent layers of the diffraction optical network is 20 cm, and the distance between the terminal reconfigurable diffraction layer and the optical detector is 15 cm.

8. The dynamic holographic generation system of claim 5, wherein, The diffraction optical network is arranged between the input plane and the optical detector by a support, and the axial alignment error between the diffraction optical networks is less than 0.1 mm.

9. The dynamic holographic generation system of claim 5, wherein, The auxiliary light source is a He-Ne laser II with a wavelength of 633 nm and a power of 1 mW.

Images