Spectral imaging method and system based on diffraction neural network
The diffraction neural network spectral imaging system solves the challenges of cophase adjustment and imaging quality in sparse aperture telescope systems, achieving efficient and high-precision spectral imaging and target recognition, and adapting to complex environments.
Patent Information
- Application Number
- CN202511485887.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-10-17
- Publication Date
- 2026-01-27
AI Technical Summary
Sparse aperture telescope systems face challenges in terms of cophase adjustment, imaging quality, and target recognition capabilities. In particular, rapid adjustment is difficult to achieve in dynamic environments. Furthermore, traditional methods suffer from high computational complexity, limited generalization ability, insufficient multi-band information fusion, and low optical computational efficiency.
A spectral imaging system based on diffraction neural networks is adopted. Through the coordinated work of the pupil remapping module, the multi-laser injection unit, the diffraction neural network module and the residual neural network module, the efficient synthesis and optical calculation of sparse aperture beams are achieved. Combined with polarization beam splitting and residual feedback mechanisms, signal processing is optimized.
It significantly improves the processing efficiency and spectral detection performance of sparse aperture beams, enhances imaging resolution, measurement accuracy, and target recognition accuracy, adapts to complex co-phase errors and scattering scenarios, and achieves efficient and high-precision spectral imaging.
Smart Images
Figure CN121409406A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of optical technology, and in particular to a spectral imaging method and system based on a diffraction neural network. Background Technology
[0002] In the field of optical imaging and target recognition, especially in applications involving large field-of-view interferometric imaging and target recognition, effectively solving the co-phase adjustment problem of sparse aperture telescopes to improve image quality and target recognition capabilities has always been a challenging issue. Traditional imaging and target recognition methods, such as deconvolution algorithms, adaptive optics, or wavefront shaping methods, can solve the image distortion problem caused by scattering media to some extent, but these methods often require a large amount of digital computation, hindering the improvement of actual frame rates, and require additional effort in downstream tasks such as image reconstruction and differentiation.
[0003] To address the aforementioned issues, this invention proposes a spectral imaging method and system based on a diffraction neural network, which realizes image reconstruction and target recognition functions in the cophase adjustment of sparse aperture telescopes. Summary of the Invention
[0004] The purpose of this application is to provide a spectral imaging method and system based on diffraction neural networks, which aims to solve the above-mentioned problems.
[0005] To achieve the above objectives, this application provides the following technical solution:
[0006] In a first aspect, this application provides a spectral imaging system based on a diffraction neural network, the system comprising: a pupil remapping module, a multi-laser injection unit, a diffraction neural network module, a beam synthesizer, and a residual neural network module;
[0007] The pupil remapping module is used to synthesize the edge sub-apertures of a large-aperture sparse aperture to generate a diffraction spot.
[0008] The multi-laser injection unit, in conjunction with the optical circulator in the residual neural network module, injects multiple laser beams as the core excitation source through the unidirectional transmission characteristics of the optical circulator, so that the multiple laser beams and the diffraction spot output by the pupil remapping module are superimposed in the optical path; wherein, when the superimposed beam propagates in the multilayer scatterer of the diffraction neural network module, a chaotic effect is generated.
[0009] The diffraction neural network module consists of a microlens array, lenses, and multilayer scatterers. It receives the diffraction spot after the chaotic effect is superimposed by the multi-laser injection unit, and completes the optical calculation of the optical signal through beam distribution of the microlens array, optical path control of the lenses, and optical modulation of the multilayer scatterers.
[0010] A beam synthesizer, located at the output end of the system, couples the optical signal processed by the diffraction neural network module to the spectrometer;
[0011] The residual neural network module includes an optical circulator, which provides an optical path for the multi-laser injection unit through the unidirectional transmission characteristics of the optical circulator; and feeds back the output signal unidirectionally to the input end to form a mixed input-output mode.
[0012] Secondly, this application provides a spectral imaging method based on a diffraction neural network, the steps of which include:
[0013] After compensating for the co-phase error of the beam at the edge of the sparse aperture, a diffraction spot is synthesized.
[0014] Feature extraction of the diffraction spot is performed using a diffraction neural network;
[0015] The generalization ability of the diffraction neural network is improved based on the principle of polarization spectroscopy.
[0016] The diffraction neural network is optimized by a residual neural network, and the output signal of the diffraction neural network is transmitted to the system input terminal.
[0017] Thirdly, this application provides a computer device, the computer device including a processor and a memory coupled to the processor, wherein the memory stores program instructions for implementing a spectral imaging method based on a diffraction neural network; the processor is used to execute the program instructions stored in the memory to implement a spectral imaging method based on a diffraction neural network.
[0018] Fourthly, this application provides a storage medium storing processor-executable program instructions for executing a spectral imaging method based on a diffraction neural network.
[0019] This application provides a spectral imaging method and system based on a diffraction neural network, which has the following advantages:
[0020] This application achieves multi-laser core injection through the cooperation of a multi-laser injection unit and an optical circulator, resulting in a chaotic effect after beam superposition. This effectively improves the randomness of photon propagation within the system and enhances its adaptability to complex co-phase errors and unknown scattering scenarios. Simultaneously, the diffraction neural network module performs high-precision optical calculations through structures such as multilayer scatterers. Combined with the unidirectional feedback mechanism of the residual neural network module, it forms an input-output hybrid mode, strengthening the cyclic optimization of the signal. Furthermore, with the efficient coupling of the optical signal by the beam synthesizer, it significantly improves the measurement accuracy and precision of spectral imaging, and overall optimizes the system's processing efficiency and spectral detection performance for sparse aperture beams. Attached Figure Description
[0021] Figure 1 This is a schematic diagram of the structure of a spectral imaging system based on a diffraction neural network according to Embodiment 1 of this application;
[0022] Figure 2 This is a schematic flowchart of a spectral imaging method based on a diffraction neural network according to Embodiment 2 of this application;
[0023] Figure 3 This is a schematic diagram of the computer device structure according to Embodiment 3 of this application;
[0024] Figure 4 This is a schematic diagram of the storage medium structure in Embodiment 4 of this application. Detailed Implementation
[0025] It should be understood that the specific embodiments described herein are merely illustrative of this application and are not intended to limit this application.
[0026] The following problems exist in the existing technology:
[0027] In the field of optical imaging, sparse aperture telescope systems achieve high-resolution imaging by combining beams from multiple sub-apertures. However, due to co-phase errors between sub-apertures (such as mechanical vibration and thermal deformation), traditional methods face the following challenges:
[0028] Phase error compensation is difficult: Existing technologies usually rely on adaptive optics or digital computing for phase correction, which involves a large amount of computation and poor real-time performance, making it difficult to meet the needs of rapid adjustment in dynamic environments.
[0029] Scattering media affect image quality: When a beam of light passes through random scatterers (such as atmospheric turbulence, biological tissue, etc.), traditional methods rely on digital deconvolution or deep learning to reconstruct the image, which has high computational complexity and limited generalization ability.
[0030] Insufficient multi-band information fusion: Existing systems typically use single-band imaging and lack joint optimization of polarization and spectral dimensions, which limits the accuracy of target recognition and classification.
[0031] Optical computing is inefficient: Traditional neural networks rely on electronic computing, while all-optical diffraction neural networks can improve speed, but they lack dynamic tunability and are difficult to adapt to different scattering conditions and task requirements.
[0032] This application first achieves pupil remapping and edge aperture beam synthesis through an adjustable focus initial coupler; then, the synthesized beam is input into a light diffraction neural network composed of a microlens array and multiple adjustable scatterers; finally, after optimization through polarization beam splitting and residual feedback, the beam is output to a spectrometer. This effectively solves the problems of difficulty in compensating for co-phase errors, interference from scattering media, and insufficient fusion of multi-band information existing in traditional technologies.
[0033] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of the embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.
[0034] Example 1
[0035] Please see Figure 1 This is a schematic diagram of the structure of a diffraction neural network-based spectral imaging system according to Embodiment 1 of this application; the specific content includes:
[0036] The pupil remapping module is used to combine the edge sub-apertures of a large-aperture sparse aperture to generate a diffraction spot.
[0037] The multi-laser injection unit, in conjunction with the optical circulator in the residual neural network module, injects multiple laser beams as the core excitation source through the unidirectional transmission characteristics of the optical circulator, so that the multiple laser beams and the diffraction spot output by the pupil remapping module are superimposed in the optical path; wherein, when the superimposed beam propagates in the multilayer scatterer of the diffraction neural network module, a chaotic effect is generated.
[0038] The diffraction neural network module consists of a microlens array, lenses, and multilayer scatterers. It receives the diffraction spot after the chaotic effect is superimposed by the multi-laser injection unit. Through beam distribution of the microlens array, optical path control of the lenses, and optical modulation of the multilayer scatterers, it completes the optical calculation of the optical signal.
[0039] A beam synthesizer, located at the output end of the system, couples the optical signal processed by the diffraction neural network module to the spectrometer;
[0040] The residual neural network module includes an optical circulator, which provides an optical path for the multi-laser injection unit through the unidirectional transmission characteristics of the optical circulator; and feeds back the output signal unidirectionally to the input end to form a mixed input-output mode.
[0041] This embodiment proposes a diffraction neural network-based spectral imaging system, aiming to overcome the limitations of traditional spectral imaging techniques in terms of resolution, imaging speed, and system complexity. By innovatively integrating multiple techniques such as pupil remapping, diffraction neural network calculation, beam combining, and residual neural network feedback, it achieves efficient and high-precision spectral imaging. The system has a compact overall structure, with each module working collaboratively to form a complete closed loop from beam acquisition and optical calculation to signal feedback optimization. Each module is described in detail below.
[0042] First, the beam enters a polarizer to filter out linearly polarized light in a single direction, eliminating interference from polarization disorder in subsequent modulation. The polarized beam then enters a vortex waveplate, where wavefront phase modulation forms a spirally distributed vortex beam. Its self-focusing properties enhance the efficiency of subsequent interactions with sparse apertures. The vortex beam then enters a multi-laser injection unit, where it is superimposed with multiple auxiliary laser beams. These laser beams are injected unidirectionally through an optical circulator in a residual neural network module to avoid back-current crosstalk. Due to differences in wavelength and phase, the superimposed beam initially possesses the basis for chaotic perturbations.
[0043] The superimposed beam is then input into the pupil remapping module to synthesize the beams from the edge sub-apertures of the large-aperture sparse aperture, correcting the optical field discontinuity caused by the sparse aperture, and ultimately generating a uniform diffraction spot with chaotic perturbations. The diffraction spot output from the pupil remapping module, calculated by the diffraction neural network optical computation, is first spatially split and focused by a microlens array, decomposing the complex spot into multiple independently processable sub-light signals. These sub-light signals undergo optical path calibration through lenses to ensure consistent propagation directions. Finally, they enter a multi-layer scatterer, where the spacing and number of scatterers are adjusted to perform phase modulation and feature extraction, completing the optical-level neural network inference calculation.
[0044] After being processed by the diffraction neural network module, a portion of the optical signal is fed back unidirectionally to the system input via an optical circulator. This circulator superimposes the initial vortex beam to form an input-output hybrid mode, enhancing the dynamic correction capability for co-phase errors. The other portion of the optical signal is fed into a beam synthesizer, which recouples the multiple sub-light signals into a unified beam, eliminating signal dispersion issues during optical calculations. The coupled beam output from the spectral detection beam synthesizer is then injected into a spectrometer to complete the acquisition and analysis of the optical signal's spectral information, achieving core functions such as co-phase error monitoring and target spectral identification.
[0045] The pupil remapping module is a crucial front-end component of this system, primarily responsible for precise beam combining of the edge sub-apertures in a large-aperture sparse aperture. In traditional large-aperture optical systems, while sparse aperture design reduces system weight and cost, the beams from the edge sub-apertures often struggle to be effectively combined directly due to propagation path and phase differences, leading to degraded image quality. This module performs pupil remapping on the edges of the large-aperture sparse aperture, combining the beams from both edge apertures to form a diffraction spot. By remapping and phase-adjusting these edge sub-aperture beams, efficient interferometric combining can be achieved in subsequent modules, resulting in a high-quality diffraction spot.
[0046] Specifically, the pupil remapping module employs a series of micro-optical components, such as microprism arrays and phase modulators. The microprism array is custom-designed based on the distribution of the edge sub-apertures and the beam characteristics. Each microprism can precisely change the beam propagation direction, guiding beams from different edge sub-apertures to a predetermined combining position. The phase modulator adjusts the beam phase in real time to compensate for phase deviations caused by optical path differences, ensuring that the beams from each sub-aperture have the same phase relationship during combining, thereby achieving optimal interference effects. After processing by the pupil remapping module, the originally dispersed edge sub-aperture beams are combined into a diffraction spot with distinct diffraction characteristics, providing an ideal input signal for subsequent diffraction neural network calculations.
[0047] In addition, before the light beam enters the pupil remapping module, it is processed by a polarizer and a vortex waveplate in sequence: the polarizer is used to filter out linearly polarized light in a single direction, providing a uniform polarization basis for the subsequent vortexization of the light beam; the vortex waveplate is used to modulate the phase of the linearly polarized light in a single direction, so that the wavefront of the light beam forms a spiral distribution, which is transformed into vortex light with self-focusing characteristics, thereby improving the energy concentration when the light beam interacts with the edge sub-aperture of the sparse aperture and reducing effective photon loss.
[0048] Furthermore, the diffraction neural network module is the core computing unit of the system, utilizing the diffraction properties of light to achieve high-speed, parallel optical computation. Compared with traditional electronic neural networks, diffraction neural networks have advantages such as high computational speed, low energy consumption, and the ability to process large amounts of data in parallel, making them particularly suitable for spectral imaging applications with high real-time requirements. This module receives the diffraction spot from the pupil remapping module and, through the light interaction between multiple scatterers, performs complex nonlinear transformations and feature extraction on the optical signal, thereby achieving preliminary analysis and processing of spectral information.
[0049] Specifically, the diffraction neural network module consists of a microlens array, lenses, and multiple scatterers. The microlens array, located at the module's input, focuses and splits the input diffraction spot into multiple sub-beams, guiding each sub-beam to a different scatterer region. Lenses further focus and collimate the beam, ensuring efficient propagation and interaction between the scatterers. The multiple scatterers are a key component of the diffraction neural network; each layer is made of a material with a specific microstructure. These microstructures produce complex scattering effects on the incident beam, altering its amplitude, phase, and polarization. By rationally designing the structure and parameters of the multiple scatterers, various optical computational functions, such as convolution and nonlinear activation, can be achieved. During beam propagation, the optical interactions between the scatterers are analogous to neuronal connections in a neural network. Through multiple scattering and interference, deep processing and feature extraction of the input optical signal are achieved. Finally, the optical signal processed by the diffraction neural network module carries rich spectral information and is transmitted to a subsequent beam synthesizer for further processing.
[0050] Furthermore, the beam synthesizer is located at the output end of the system. Its main function is to efficiently couple the multiple optical signals processed by the diffraction neural network module and guide them to the spectrometer for final spectral analysis. Since the optical signals processed by the diffraction neural network module are distributed at different spatial positions and angles, directly inputting them into the spectrometer would lead to signal loss and a decrease in imaging quality. The beam synthesizer, through precise optical design, recombines these dispersed beams into a single beam with specific direction and wavefront characteristics, ensuring that the optical signal can efficiently enter the spectrometer, thereby improving the sensitivity and resolution of spectral imaging.
[0051] Specifically, the beam synthesizer employs a series of optical elements, such as mirrors, prisms, and coupling lenses. Mirrors and prisms are used to change the propagation direction of the beam, guiding beams from different paths to the same convergence point. The coupling lens further focuses and collimates the converged beam, ensuring accurate coupling to the spectrometer's input port. During beam synthesis, precise control and optimization of the beam can be achieved by accurately adjusting the position and angle of each optical element, ensuring the synthesized beam has optimal quality and coupling efficiency. Furthermore, the beam synthesizer also possesses beam shaping capabilities, allowing adjustment of the beam's shape and size according to the spectrometer's requirements, further enhancing system compatibility and performance.
[0052] Furthermore, the residual neural network module introduces the concept of residual learning. By unidirectionally feeding the output signal back to the input, a hybrid input-output mode is formed, thereby achieving continuous optimization and improvement of system performance. In traditional spectral imaging systems, due to manufacturing errors of optical components, environmental interference, and other factors, the system output signal often contains certain errors and noise. The residual neural network module performs feedback analysis on the output signal, extracts the error information, and feeds it back to the input. This information is then superimposed and corrected with the original input signal, thereby achieving real-time compensation and correction of system errors. This feedback mechanism enables the system to continuously learn and adapt to environmental changes, improving system stability and imaging accuracy.
[0053] The residual neural network module includes key components such as an optical circulator. An optical circulator is an optical device with unidirectional light transmission characteristics, enabling unidirectional cyclic transmission of optical signals and ensuring that the output signal is accurately fed back to the input without interfering with the original input signal. In the system, the optical circulator guides the output signal, analyzed by the spectrometer, to the feedback path, where it is preprocessed and adjusted by a series of optical processing components, such as filters and amplifiers. The preprocessed feedback signal is then fed back to the input of the pupil remapping module, where it is superimposed and mixed with the original input beam. In this way, the system can sense output errors in real time and dynamically adjust the input signal using the feedback mechanism of the residual neural network, thereby optimizing and correcting the entire spectral imaging system. Furthermore, the residual neural network module can work collaboratively with the diffraction neural network module, using feedback information to adjust and optimize the parameters of the diffraction neural network online, further improving the system's computational performance and imaging quality.
[0054] In summary, the diffraction neural network-based spectral imaging system proposed in this application, through the collaborative work of the pupil remapping module and the diffraction neural network module, enables efficient synthesis and complex optical calculations of large-aperture sparse-aperture beams, significantly improving the system's imaging resolution and processing speed. Specifically, the diffraction neural network utilizes the diffraction properties of light to achieve parallel computation, exhibiting lower energy consumption and higher computational efficiency compared to traditional electronic neural networks, thus meeting the requirements of real-time spectral imaging. Furthermore, the residual neural network module achieves real-time compensation and correction of system errors through a feedback mechanism, enabling the system to adaptively respond to environmental changes and optical component errors, thereby improving system stability and imaging accuracy.
[0055] Example 2
[0056] Please see Figure 2 This is a flowchart illustrating a spectral imaging method based on a diffraction neural network according to Embodiment 2 of this application; the steps include:
[0057] S1: Synthesize the diffraction spot after compensating for the phase error of the beam at the edge of the sparse aperture.
[0058] In this embodiment, a pupil remapping technique is employed to address the optical path difference problem of the edge sub-apertures. Specifically, a pupil remapping element, such as a microprism array or a deformable mirror, is placed in the outgoing optical path of each edge sub-aperture. The microprism array consists of multiple tiny prisms, each of which can be customized according to the beam characteristics of the edge sub-aperture, precisely changing the beam propagation direction and guiding beams from different edge sub-apertures to a predetermined combining position. The deformable mirror, through the independent control of multiple tiny driving units on its surface, adjusts the wavefront phase of the beam in real time to compensate for the phase deviation caused by the optical path difference. Through the synergistic effect of the pupil remapping elements, the beams of the edge sub-apertures are spatially redistributed and phase-adjusted, creating favorable conditions for subsequent beam combining.
[0059] After pupil remapping, the edge sub-aperture beams are guided to the synthesis region. Due to the identical phase relationship and suitable spatial distribution of each beam, interference occurs between them, forming diffraction spots with clear diffraction fringes. These diffraction spots contain rich spectral information, and their diffraction characteristics are closely related to parameters such as the wavelength and phase of the input beam, providing ideal input signals for subsequent diffraction neural network processing.
[0060] S2: Extract features from the diffraction spot using a diffraction neural network.
[0061] In this embodiment, the diffraction neural network is a novel optical computing system that realizes information processing based on the diffraction and interference properties of light. Specifically, it is composed of a microlens array, lenses, and a multilayer scatterer. The microlens array is located at the system's input end, and its function is to focus and split the input diffracted light spot into multiple sub-beams, guiding them to different scatterer regions. The lenses are used to further focus and collimate the beams, ensuring that the beams can propagate and interact efficiently between the scatterers. The multilayer scatterer is the core component of the diffraction neural network. Each scatterer layer is made of a material with a specific microstructure, which can produce complex scattering effects on the incident beam, changing its amplitude, phase, and polarization characteristics.
[0062] Furthermore, to address different levels of co-phase error, a neural network structure was constructed by adjusting the spacing and number of scatterers. The spacing and number of scatterers directly affect the propagation path and interaction mode of the light beam between scatterers, thus determining the computational power and feature extraction capability of the neural network. When the co-phase error is large, appropriately increasing the number of scatterers and adjusting their spacing can enhance the scattering and interference effects of the light beam, improving the neural network's tolerance and correction capability to errors; when the co-phase error is small, optimizing the scatterer parameters can enable the neural network to extract spectral feature information more efficiently.
[0063] To achieve precise control and optimization of the diffraction neural network, this application employs an adjustable-focus initial coupler and an initial reflection angle adjustment technique for network parameter setting. The adjustable-focus initial coupler can adjust the focusing position and coupling efficiency of the input beam in real time according to the characteristics of the input beam and system requirements, ensuring that the beam accurately enters the diffraction neural network. By adjusting the initial reflection angle, the propagation direction and interference conditions of the beam in the system can be changed, further optimizing the performance of the neural network. Simultaneously, combined with numerical simulation methods, the system output under different parameter combinations is calculated and analyzed to determine the optimal network parameter settings, enabling the diffraction neural network to achieve efficient and stable spectral information processing under various operating conditions.
[0064] S3: Improve the generalization ability of the diffraction neural network based on the principle of polarization spectroscopy.
[0065] In this embodiment, the principle of polarization beam splitting is utilized to enhance the generalization ability of the diffraction neural network. Polarization is an important characteristic of light; light of different wavelengths has different polarization states during propagation. Utilizing the principle of polarization beam splitting, a polarization beam splitter, such as a polarization beam splitter, is introduced at the output end of the diffraction neural network. The polarization beam splitter can separate the beam processed by the diffraction neural network according to its polarization direction, forming two or more independent polarization channels. The beam in each polarization channel has relatively singular polarization characteristics, reducing mutual interference between light with different polarization states, thereby improving the diffraction neural network's ability to extract and generalize different spectral features.
[0066] Furthermore, this application employs a combination of filters with different wavelengths to adjust the wavelengths used for system training. By selecting appropriate filters, light signals of specific wavelengths can be filtered out for analysis by the spectrometer, while simultaneously providing targeted training for the diffraction neural network. The interaction characteristics between the beam and the scattering body, as well as the polarization-splitting effect, differ across wavelengths. Through multi-band training, the diffraction neural network can learn richer spectral feature information, improving its adaptability and resolution in different spectral scenarios.
[0067] S4: Optimize the diffraction neural network using a residual neural network, and transmit the output signal of the diffraction neural network to the system input terminal.
[0068] In this embodiment, the residual neural network is a deep learning model that solves the gradient vanishing problem that may occur during the training process of traditional deep neural networks by introducing residual connections. It can learn the residual information between the input and output, thereby optimizing and improving system performance. Specifically, this application utilizes an optical circulator to construct an optical residual neural network. An optical circulator is an optical device with unidirectional light transmission characteristics, enabling unidirectional cyclic transmission of optical signals. By introducing the optical circulator into the diffraction neural network, the system's output signal is unidirectionally transmitted to the input end, superimposed with the original input signal to form a mixed input-output mode, thus constructing the residual neural network structure.
[0069] In the residual neural network structure, the output signal is fed back to the input through an optical circulator, and together with the original input signal, it enters the diffraction neural network system for further processing. This feedback mechanism enables the system to perceive output errors in real time and dynamically correct and optimize the input signal through the learning and adjustment capabilities of the residual neural network. After multiple feedback iterations, the optimized signal is finally input into the spectrometer for spectral analysis. Due to the feedback optimization effect of the residual neural network, the signal quality input to the spectrometer is significantly improved, resulting in more accurate and clearer spectral images, providing a reliable guarantee for the extraction and analysis of spectral information.
[0070] In summary, this application first achieves pupil remapping and edge aperture beam synthesis through an adjustable focus initial coupler, then inputs the synthesized beam into a light diffraction neural network composed of a microlens array and a multilayer adjustable scatterer; finally, after optimization through polarization beam splitting and residual feedback, the output is sent to a spectrometer. This effectively solves the problems of difficult co-phase error compensation, scattering medium interference, and insufficient multi-band information fusion in traditional technologies, achieving efficient and high-precision spectral imaging and providing strong support for research and applications in related fields.
[0071] Example 3
[0072] Please see Figure 3 This is a schematic diagram of the computer device structure according to Embodiment 3 of this application. The computer device 50 includes a processor 51 and a memory 52 coupled to the processor 51.
[0073] The memory 52 stores program instructions for implementing the above-described spectral imaging method based on a diffraction neural network.
[0074] The processor 51 is used to execute program instructions stored in the memory 52 to implement a spectral imaging based on a diffraction neural network.
[0075] The processor 51 can also be referred to as a CPU (Central Processing Unit).
[0076] Processor 51 may be an integrated circuit chip with signal processing capabilities. Processor 51 may also be a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), an off-the-shelf programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. A general-purpose processor may be a microprocessor or any conventional processor.
[0077] Example 4
[0078] Please see Figure 4 This is a schematic diagram of the storage medium in Embodiment 4 of this application. The storage medium in this embodiment stores a program file 61 capable of implementing all the above methods. This program file 61 can be stored in the storage medium in the form of a software product, including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) or processor to execute all or part of the steps of the methods of various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks, or devices such as computers, servers, mobile phones, and tablets.
[0079] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, apparatus, article, or method that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, apparatus, article, or method. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, apparatus, article, or method that includes that element.
[0080] The above description is only a preferred embodiment of this application and does not limit the patent scope of this application. Any equivalent structural or procedural changes made based on the content of this application's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.
[0081] Although embodiments of this application have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of this application, the scope of which is defined by the appended claims and their equivalents.
[0082] Of course, the present invention may have many other embodiments. Based on this embodiment, other embodiments obtained by those skilled in the art without any creative effort are all within the scope of protection of the present invention.
Claims
1. A spectral imaging system based on a diffraction neural network, characterized in that, The system includes: a pupil remapping module, a multi-laser injection unit, a diffraction neural network module, a beam synthesizer, and a residual neural network module; The pupil remapping module is used to synthesize the edge sub-apertures of a large-aperture sparse aperture to generate a diffraction spot. The multi-laser injection unit, in conjunction with the optical circulator in the residual neural network module, injects multiple laser beams as the core excitation source through the unidirectional transmission characteristics of the optical circulator, so that the multiple laser beams and the diffraction spot output by the pupil remapping module are superimposed in the optical path; wherein, when the superimposed beam propagates in the multilayer scatterer of the diffraction neural network module, a chaotic effect is generated. The diffraction neural network module consists of a microlens array, lenses, and multilayer scatterers. It receives the diffraction spot after the chaotic effect is superimposed by the multi-laser injection unit, and completes the optical calculation of the optical signal through beam distribution of the microlens array, optical path control of the lenses, and optical modulation of the multilayer scatterers. The beam synthesizer, located at the output end of the system, couples the optical signal processed by the diffraction neural network module to the spectrometer; The residual neural network module includes an optical circulator, which provides an optical path for the multi-laser injection unit through the unidirectional transmission characteristics of the optical circulator; and feeds back the output signal unidirectionally to the input end to form a mixed input-output mode.
2. The spectral imaging system based on a diffraction neural network according to claim 1, characterized in that, The pupil remapping module achieves phase matching of the edge aperture by adjusting the reflection angle of the initial coupler, which includes a micro-optical structure with adjustable numerical aperture.
3. The spectral imaging system based on a diffraction neural network according to claim 1, characterized in that, The diffraction neural network module specifically includes: dynamically adjusting the axial spacing between each scatterer; changing the number of layers of scatterers; and using a piezoelectric actuator to correct co-phase error in real time.
4. The spectral imaging system based on a diffraction neural network according to claim 1, characterized in that, The pupil remapping module also includes a polarizer and a vortex wave plate; Before the light beam is incident on the pupil remapping module, it is processed by a polarizer and a vortex waveplate in sequence. The polarizer is used to filter out linearly polarized light in a single direction; the vortex waveplate is used to modulate the phase of the linearly polarized light in a single direction, so that the wavefront of the light beam forms a spiral distribution and is transformed into vortex light with self-focusing characteristics.
5. A spectral imaging method based on a diffraction neural network, characterized in that, include: After compensating for the co-phase error of the beam at the edge of the sparse aperture, a diffraction spot is synthesized. Feature extraction of the diffraction spot is performed using a diffraction neural network; The generalization ability of the diffraction neural network is improved based on the principle of polarization spectroscopy. The diffraction neural network is optimized by a residual neural network, and the output signal of the diffraction neural network is transmitted to the system input terminal.
6. A computer device, characterized in that, The computer device includes a processor and a memory coupled to the processor, wherein the memory stores program instructions for implementing the spectral imaging method based on a diffraction neural network as described in claim 5; the processor is used to execute the program instructions stored in the memory to implement a spectral imaging method based on a diffraction neural network.
7. A storage medium, characterized in that, It stores processor-executable program instructions for performing the spectral imaging method based on a diffraction neural network as described in claim 5.