Speckle identification and deep learning-based spectral measurement method and calculation reconstruction spectrometer

Abstract

The invention provides a spectral measurement method based on speckle identification and deep learning and a calculation reconstruction spectrograph, and the method comprises the steps: inputting speckle image time sequence data into a deep learning model, so as to output a wavelength time sequence change rule of incident light to be measured corresponding to the speckle image time sequence data, the speckle image time series data is generated by excitation of incident light to be measured under any incident vortex light order condition, and the deep learning model comprises a non-convolutional neural network model; based on the speckle image time sequence data, determining a light intensity time sequence change rule of incident light to be measured corresponding to the speckle image time sequence data; and reconstructing the spectrum of the incident light to be measured based on the wavelength time sequence change rule and the light intensity time sequence change rule. According to the invention, the time sequence change rule and the speckle image details of the speckle image data acquired for a long time can be captured, so that the spectrum restoration precision of the incident light to be measured is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of deep learning and spectral measurement, and particularly relates to a spectral measurement method and a computational reconstruction spectrometer based on speckle identification and deep learning. BACKGROUND

[0002] Spectral analysis has important application value in the fields of chemical detection, space exploration, etc., and thus the development of high-performance spectrometers has significant practical significance. According to the working principle, traditional spectrometers include dispersive spectrometers, narrow-band filter spectrometers, and Fourier transform spectrometers. The dispersive spectrometer is usually large in size due to the limitation of traditional dispersive elements such as diffraction gratings. The narrow-band filter spectrometer has a complex process. The Fourier transform spectrometer using an on-chip integrated system faces the problems of high process requirements and optical path limitations on a silicon substrate.

[0003] In order to solve the inherent design defects of traditional spectrometers, a new type of spectrometer that assists the optical system through algorithms, i.e., a computational reconstruction spectrometer, has appeared in recent years. The computational reconstruction spectrometer can complement the inherent defects of the optical system through the cooperative optimization of the optical system and the algorithm, improve the theoretical performance of the spectrometer, realize small size, high precision, low cost, and high robustness of spectral detection, and break through the limitations of traditional spectrometers. There are currently various optical system design and algorithm implementation schemes for reconstructing spectra. One type of computational reconstruction spectrometer uses a random scattering medium as a dispersive element, and designs an algorithm to restore the mapping relationship between the exit speckle and the incident wavelength and intensity based on the wavelength sensitivity of the exit speckle formed by the coherent light passing through the scattering medium. In practical applications, the random scattering medium used to generate the exit speckle includes photonic crystals, thin quartz lenses, tapered optical fibers, and spiral waveguides, and the reconstruction algorithm usually restores the scattering medium transmission matrix pseudo-inverse by calibrating the speckle or uses a deep learning model to identify and classify the speckle images of different wavelengths and intensities excited by the coherent light source.

[0004] In existing computational reconstruction spectrometers based on identification of scattered medium exit light field, the scattered medium as a dispersive element has problems such as complex preparation process and high preparation cost, and in the design of identification algorithm, the pseudo-inverse method of transfer matrix is adopted, which has high requirements for system stability, and the exit light field detection method is complex, which requires obtaining light field information in a short time. In view of the above problems, the existing research proposes a spectral reconstruction solution based on a multimode optical fiber and deep learning. The specific implementation of the solution is as follows: a multimode optical fiber is used as a dispersive element, and a charge-coupled device camera is used to collect and record speckle images at the exit end of the optical fiber under static conditions, and then a deep learning model is used to realize classification and identification of small sample speckle images. On the one hand, this solution makes full use of the characteristics of the multimode optical fiber, such as low loss, low cost, easy coupling and miniaturization; on the other hand, it takes advantage of the advantages of convolutional neural networks in small sample image processing, including low memory occupation and training cost, and a simplified algorithm implementation and sample collection method compared with the transfer matrix method. However, the solution has the following shortcomings: 1) Convolutional neural networks have limitations in capturing temporal variation and image details of speckle images collected over a long period of time, resulting in limited spectral reconstruction accuracy; 2) The above speckle image collection method cannot effectively suppress the intermodal coupling effect caused by bending of the multimode optical fiber, which can cause distortion of speckle image features and decrease in wavelength recognition accuracy, and has poor robustness under fiber deformation conditions; 3) Due to the physical limitation of the width of spectral correlation, the number of effective speckle image samples available for model training is fixed, which limits the performance optimization space of the deep learning model. SUMMARY

[0005] In view of this, the embodiments of the present application provide a speckle identification and deep learning based spectral measurement method and a computational reconstruction spectrometer to eliminate or improve one or more defects in the prior art.

[0006] One aspect of the present application provides a speckle identification and deep learning based spectral measurement method, which comprises the following steps: inputting speckle image time series data into a deep learning model to output wavelength time variation of the speckle image time series data corresponding to the measured incident light, wherein the speckle image time series data is generated by exciting the measured incident light under any incident vortex light order condition, and the deep learning model comprises a non-convolutional neural network model; determining the light intensity time variation of the speckle image time series data corresponding to the measured incident light based on the speckle image time series data; and reconstructing the spectrum of the measured incident light based on the wavelength time variation and the light intensity time variation.

[0007] In some embodiments of the present application, the deep learning model is pre-trained by the following steps: training the preset deep learning model based on a plurality of speckle image data with respective incident light wavelength labels, so that the trained deep learning model can output the wavelength of the incident light corresponding to the speckle image data based on the speckle image data, wherein the plurality of speckle image data is generated by exciting a plurality of wavelengths of incident light under a plurality of incident vortex light order conditions, and each speckle image data is an average of a plurality of speckle image data generated by exciting each wavelength of incident light under a plurality of fiber deformation conditions corresponding to each incident vortex light order.

[0008] In some embodiments of the present application, before the step of training the preset deep learning model based on a plurality of speckle image data with respective incident light wavelength labels, the training process further comprises: preprocessing a plurality of speckle image data to obtain a plurality of preprocessed speckle image data, the preprocessing including removing black parts of the speckle image data.

[0009] In some embodiments of the present application, the plurality of wavelengths of incident light is obtained by multiple sampling, and the wavelength sampling interval is determined by estimating the minimum spectral resolution of the computational reconstruction spectrometer.

[0010] In some embodiments of the present application, the non-convolutional neural network model comprises a visual model based on a Transformer architecture, and the visual model comprises a Swin Transformer model.

[0011] Another aspect of the present application provides a computational reconstruction spectrometer, comprising: a tunable laser, a first polarizer, a spatial light modulator, a second polarizer, a first fiber collimator, a multimode fiber, a second fiber collimator, a near-infrared charge-coupled device camera, and a computer device; The tunable laser is used to provide incident light. The first polarizer is used to modulate the incident light into polarized light of a specific polarization state, outputting circularly polarized light. The spatial light modulator is used to load the circularly polarized light with a controllable order spiral wavefront, outputting circularly polarized light under an arbitrary incident vortex light order condition. The second polarizer is used to modulate the circularly polarized light output by the spatial light modulator into polarized light of a specific polarization state, outputting circularly polarized light. The first fiber collimator is used to transmit the circularly polarized light output by the second polarizer to the multimode fiber. The multi-mode optical fiber is used to generate speckles at an output end of the second fiber collimator based on the circularly polarized light output by the first fiber collimator, and the second fiber collimator is used to convert the divergent light transmitted by the multi-mode optical fiber into collimated light; The near-infrared charge-coupled device camera is used to receive the speckles and record and output speckle image data. The computer device includes a processor and a memory, and the memory stores computer instructions, and the processor is configured to execute the computer instructions stored in the memory, and when the computer instructions are executed by the processor, the spectrometer implements the steps of the aforementioned speckle identification and deep learning based spectral measurement method.

[0012] In some embodiments of the present application, the spectrometer further comprises a fiber coupler and a manual two-dimensional translation stage, the fiber coupler is used to couple the circularly polarized light output by the first fiber collimator to the multi-mode optical fiber, and the manual two-dimensional translation stage is used to move the multi-mode optical fiber in the x-axis or y-axis direction according to a plurality of preset fiber displacement points, so that the multi-mode optical fiber generates different degrees of deformation and generates speckles under a plurality of fiber deformation conditions, so that the near-infrared charge-coupled device camera records and outputs speckle image data under a plurality of fiber deformation conditions.

[0013] In some embodiments of the present application, the multi-mode optical fiber comprises a ring-core multi-mode optical fiber.

[0014] Another aspect of the present application provides a computer readable storage medium having a computer program stored thereon, the computer program being executed by a processor to implement the steps of the aforementioned speckle identification and deep learning based spectral measurement method.

[0015] Another aspect of the present application provides a computer program product comprising computer instructions, the computer instructions being executed by a processor to implement the steps of the aforementioned speckle identification and deep learning based spectral measurement method.

[0016] The speckle identification and deep learning based spectral measurement method and the computer reconstructed spectrometer of the present application can capture the speckle image data time sequence variation law and speckle image details when long-time acquisition is performed, thereby improving the spectral recovery accuracy of the to-be-measured incident light.

[0017] Additional advantages, objects, and features of the application will be set forth in part by the description that follows, and will in part be apparent to those of ordinary skill in the art upon examination of the following or can be learned from practice of the application. The objects and other advantages of the application can be realized and attained by the structure particularly pointed out in the specification as well as in the appended drawings.

[0018] Those skilled in the art will understand that the objects and advantages of the application can be realized and attained by means of the application especially pointed out in the following detailed description. BRIEF DESCRIPTION OF DRAWINGS

[0019] The accompanying drawings, which are included to provide a further understanding of the application and are incorporated in and constitute a part of this application, illustrate embodiments of the application and together with the detailed description serve to explain the principles of the application.

[0020] Figure 1 A flowchart of a spectral measurement method based on speckle recognition and deep learning in an embodiment of the application; Figure 2 A specific flowchart of the steps of training and testing a deep learning model in an embodiment of the application; Figure 3 A structural diagram of a computational reconstruction spectrometer in an embodiment of the application. DETAILED DESCRIPTION

[0021] To make the objects, technical solutions and advantages of the application clearer, further detailed description of the application will be given below in conjunction with the embodiments and drawings. Here, the illustrative embodiments of the application and their descriptions are used to explain the application but are not intended to limit the application.

[0022] It should also be noted that, in order to avoid obscuring the application due to unnecessary details, only structures and / or processing steps closely related to the solutions according to the application are shown in the drawings, and other details not closely related to the application are omitted.

[0023] It should be emphasized that the term “comprises / comprising” when used in this text indicates the presence of the stated features, elements, steps or components but does not exclude the presence or addition of one or more other features, elements, steps or components.

[0024] It should also be noted that, if not specifically stated, the term “connected” in this text can not only mean direct connection but also indirect connection with an intermediate.

[0025] In the following, embodiments of the application will be described with reference to the drawings. In the drawings, the same reference numerals represent the same or similar parts or the same or similar steps.

[0026] To overcome the defects in the prior art, the embodiments of the application propose a spectral measurement method based on speckle recognition and deep learning and a computational reconstruction spectrometer, which can capture the time sequence variation law and speckle image details of long-time collected speckle image data, thereby improving the spectral recovery accuracy of the to-be-measured incident light.

[0027] Figure 1This is a schematic flowchart of a spectral measurement method based on speckle recognition and deep learning in one embodiment of the present invention. Figure 1 As shown, the method includes the following steps: Step S110: Input the speckle image time series data into the deep learning model to output the wavelength time series variation law of the incident light to be tested corresponding to the speckle image time series data. The speckle image time series data is generated by the incident light to be tested under arbitrary incident vortex order conditions. The deep learning model includes a non-convolutional neural network model.

[0028] Specifically, speckle image time-series data is formed by generating multiple speckle images at different times through continuous emission of coherent light over time. Since speckle images possess rich local details and their time-series data implicitly contains dynamic changes in the transfer matrix, an efficient deep learning model is needed for feature extraction. Compared to traditional convolutional neural networks, this method employs a non-convolutional neural network model to more effectively capture the local features of speckle images and performs better in long-sequence image classification tasks. By inputting the speckle image time-series data into the non-convolutional neural network model, the model predicts the temporal variation of the wavelength of the incident light to be measured, thereby reconstructing the spectrum of the incident light and improving spectral accuracy.

[0029] Step S120: Determine the temporal variation law of the light intensity of the incident light to be measured corresponding to the speckle image temporal data based on the speckle image temporal data.

[0030] Specifically, the intensity of the incident light under test at different times is determined by the gray values ​​of each speckle image in the speckle image time series data, thereby determining the temporal variation law of the intensity and reconstructing the spectrum of the incident light under test.

[0031] Step S130: The spectrum of the incident light to be measured is reconstructed based on the wavelength time-series variation law and the light intensity time-series variation law.

[0032] Specifically, the spectrum of the incident light obtained in this step includes the variation of the light intensity at different wavelengths, which can be represented as a spectrum with wavelength on the horizontal axis and light intensity on the vertical axis. The reconstructed continuous spectrum of the incident light can then be derived.

[0033] In some embodiments, the deep learning model in step S110 is pre-trained through the following steps: training a preset deep learning model based on multiple speckle image data with their own incident light wavelength labels, so that the trained deep learning model can output the wavelength of the incident light corresponding to the speckle image data based on the speckle image data, wherein the multiple speckle image data are generated by the excitation of multiple wavelengths of incident light under multiple incident vortex light order conditions, and each speckle image data is the average value of multiple speckle image data generated by the excitation of multiple fiber deformation conditions corresponding to each incident vortex light order.

[0034] Specifically, the dataset used to train the deep learning model includes multiple speckle image data collected and their corresponding multiple incident light wavelength labels. Figure 2 This is a schematic diagram illustrating the specific process of deep learning model training and testing in one embodiment of the present invention, as shown below. Figure 2 As shown, during training, the data in this dataset can be divided into training, validation, and test sets at proportions of 80%, 10%, and 10%, respectively. The training set is used to train the deep learning model, the validation set is used to fine-tune the model's parameters during training, and the test set is used to analyze the accuracy of the incident light wavelength prediction results by comparing the predicted incident light wavelength of the speckle image data with the corresponding incident light wavelength label, thereby evaluating and validating the performance of the deep learning model. Through multiple iterations of training and optimization, the trained deep learning model is obtained. By acquiring speckle images of optical fibers excited by incident vortex light of different orders under different optical fiber deformation perturbations, and averaging the speckle images under different optical fiber deformations, training data augmentation for deep learning model training is achieved, significantly improving the speckle feature extraction capability of the deep learning model. At the same time, the trained deep learning model can overcome the influence of optical fiber bending and environmental temperature changes on the spectral reconstruction accuracy, and can improve the recognition accuracy of speckle images corresponding to the incident light wavelength under slight optical fiber deformation perturbations, thereby making the model and system more robust and the spectral reconstruction accuracy higher.

[0035] In some embodiments, prior to the step of training a preset deep learning model based on multiple speckle image data each labeled with its own incident light wavelength, the training process further includes the following steps: Multiple speckle image data are preprocessed to obtain multiple preprocessed speckle image data, wherein the preprocessing includes removing the black parts in the speckle image data.

[0036] Specifically, such as Figure 2As shown, after acquiring the dataset for model training, the speckle image data in the dataset is preprocessed. Specifically, the extra black parts of each speckle image data are removed by cropping, leaving only the main body of the image data, thus removing the interference of the training data on the model.

[0037] In some embodiments, the incident light of the plurality of wavelengths is obtained by multiple samplings, and the wavelength sampling interval is determined by estimating the minimum spectral resolution of the calculated reconstructed spectrometer.

[0038] Specifically, assume that the operating wavelength range of the reconstructed spectrometer in practical applications is (λ1, λ2), the wavelength sampling interval is Δλ, the speckle image generated by excitation at wavelengths λ∈(λ1, λ2) of coherent light is S(λ), and the light field intensity of the excitation speckle is I(λ). Also define the spectral correlation width as δλ and the spectral correlation coefficient as... Zero wavelength offset autocorrelation value Since the grayscale value of the speckle image is proportional to the light field intensity captured by the near-infrared charge-coupled device camera in the computational reconstructed spectrometer, the light intensity of the excitation source can be obtained by normalizing the grayscale value. Specifically, this method uses a mean-variance normalization method. The grayscale data of the speckle image is converted into a light intensity distribution with a mean of 0 and a standard deviation of 1. When designing the spectrometer, the value of the spectral correlation width is estimated and verified using the following method. Based on the verification results, the estimated value of the spectral correlation width is adjusted and verified again until the condition is met: when the input wavelength changes by δλ, the output speckle correlation coefficient drops to half of its initial value, i.e. When the speckle pattern is essentially uncorrelated, δλ represents the smallest wavelength difference that the speckle imager can resolve. The wavelength sampling interval of the speckle imager or the speckle image sampling interval Δλ should be greater than or equal to the calculated theoretical minimum spectral resolution. For multiple speckle images obtained from multiple samplings, the spectral correlation width under all wavelength offset starting conditions is calculated frame by frame. By calculating the average of all calculated spectral correlation widths, the theoretical spectral correlation width within the spectrometer's operating band is obtained.

[0039] In some embodiments, the non-convolutional neural network model includes a visual model based on the Transformer architecture, wherein the visual model includes the Swing Transformer model.

[0040] Compared to traditional convolutional neural networks, the window attention mechanism of the Swin Transformer model can more effectively capture local features of images and performs better in long sequence image classification tasks. Therefore, in this embodiment, as Figure 2As shown, using the Swing Transformer model to complete the image classification task of fiber speckle can improve the accuracy of speckle recognition, thereby improving the accuracy and robustness of spectral reconstruction.

[0041] Corresponding to the above method, embodiments of the present invention also provide a computational reconstruction spectrometer. Figure 3 This is a schematic diagram of the structure of a computational reconstruction spectrometer in one embodiment of the present invention, as shown below. Figure 3 As shown, the spectrometer includes: a tunable laser, a first polarizer, a spatial light modulator, a second polarizer, a first fiber collimator, a multimode fiber, a second fiber collimator, a near-infrared charge-coupled device camera, and a computer device. The tunable laser is used to provide incident light; The first polarizer is used to modulate the incident light into polarized light with a specific polarization state, and output circularly polarized light; The spatial light modulator is used to load a controllable order spiral wavefront onto the circularly polarized light, and output circularly polarized light under arbitrary incident vortex light order conditions. The second polarizer is used to modulate the circularly polarized light output by the spatial light modulator into polarized light with a specific polarization state, and output circularly polarized light; The first fiber collimator is used to transmit the circularly polarized light output from the second polarizer to the multimode fiber; The multimode fiber is used to generate speckle at the output end of the second fiber collimator based on the circularly polarized light output from the first fiber collimator using mode interference characteristics. The second fiber collimator is used to convert the divergent light transmitted by the multimode fiber into collimated light. The near-infrared charge-coupled device camera is used to receive the speckle and record and output speckle image data; The computer device includes a processor and a memory. The memory stores computer instructions, and the processor executes the computer instructions stored in the memory. When the computer instructions are executed by the processor, the spectrometer implements the steps of the speckle recognition and deep learning-based spectral measurement method described in any of the foregoing embodiments.

[0042] Specifically, the computational reconstruction spectrometer designed in this invention mainly operates in a low-power, wide-pulse light source environment. Therefore, the nonlinear transmission effect caused by mode coupling in multimode fiber is not considered in the design, and the multimode fiber is approximately assumed to satisfy linear effects. The tunable laser, as a coherent light source, emits coherent incident light. When the incident light propagates within the multimode fiber, the excited guided mode phase satisfies φ = β(λ)L, where the transmission constant β corresponds to different transmission modes in the multimode fiber, λ represents the wavelength of the incident light, and L represents the length of the multimode fiber. At the output end of the multimode fiber, the light fields excited by different modes with different phases superimpose, causing interference. Furthermore, because multimode fiber supports a wider distribution of modes, mode coupling also exists between different modes propagating independently within the fiber. Therefore, under the combined effect of mode interference and mode coupling, randomly distributed dark and bright spots, i.e., fiber speckle, are formed at the output end of the multimode fiber. Speckle patterns exhibit significant wavelength dependence when the multimode fiber length is fixed. The light intensity follows a negative exponential statistical property, with significant intensity fluctuations. This makes the resulting speckle pattern sensitive to changes in the incident light wavelength and possesses rich image details. Based on these characteristics, the incident light spectrum can be reconstructed by calibrating the speckle patterns generated by incident light excited at different wavelengths. Furthermore, in this computational reconstructed spectrometer, the spatial light modulator can be a liquid crystal phase spatial light modulator or a phase-type liquid crystal spatial light modulator.

[0043] In some embodiments, the spectrometer further includes a manual two-dimensional translation stage, which is used to move the multimode fiber in parallel along the x-axis or y-axis direction according to multiple preset fiber displacement points, so that the multimode fiber undergoes different degrees of deformation and generates speckle under multiple fiber deformation conditions, so that the near-infrared charge-coupled device camera can record and output speckle image data under multiple fiber deformation conditions.

[0044] Specifically, a manual two-dimensional translation stage can perform multiple parallel movements of a multimode fiber along the x-axis or y-axis with an accuracy of 1mm. Each movement of the multimode fiber to a preset fiber displacement point results in one deformation, with varying degrees of deformation at different displacement points. By pre-selecting multiple fiber displacement points on the manual two-dimensional translation stage and moving the multimode fiber to these points along the x-axis or y-axis, speckle image data under different fiber deformation perturbation conditions can be acquired, thereby constructing a dataset for training a deep learning model.

[0045] Specifically, such as Figure 2As shown, the minimum spectral resolution of the reconstructed spectrometer is first estimated to determine the appropriate sampling interval Δλ. Then, keeping the order of any incident vortex light of the spatial light modulator constant, with the tunable laser on, a fixed multimode fiber is moved to different fiber displacement points using a manual two-dimensional translation stage, causing the fiber to deform to varying degrees. Within the spectrometer's operating wavelength range (λ1, λ2), multiple wavelengths of incident light are sampled at a sampling interval Δλ that satisfies the spectral resolution δλ. A near-infrared charge-coupled device (CCPD) camera captures speckle images at the same time interval as the optical switch setting of the tunable laser on the camera element. Assuming m fiber displacement points are taken, the number of speckle images that can be acquired at each displacement point is... The total number of speckle images that can be acquired at all displacement points is Next, the average of all speckle images under all fiber deformation perturbation conditions at any fixed wavelength λ is calculated, and the resulting average speckle image is labeled as the fiber speckle image generated under that wavelength condition. In other words, the number of speckle image data that can be acquired under this incident vortex light order is... The specific calculation method for the average value of all speckle images measured at m fiber displacement points under a given wavelength λ is as follows: Calculate the average value of all speckle images at the same pixel position to obtain an average speckle image. This method can remove noise from the speckle image and extract its static features. Then, the order of the incident vortex light of the spatial light modulator is changed multiple times, and the above steps are repeated each time the order is changed, thereby constructing a dataset for training a deep learning model. Let n different orders of incident vortex light be obtained under m different fiber deformations to represent speckle patterns, i.e., the number of speckle images that can be obtained is... The final number of average speckle images obtained after averaging is: N′ represents the number of speckle images in the dataset used to train the deep learning model. By averaging speckle images under different fiber deformations, the accuracy of identifying the incident light wavelength in speckle images under slight fiber deformation perturbations can be improved, thereby enhancing the robustness of the model and system. This speckle image data acquisition method addresses the problem in existing technologies where the physical limitation of spectral correlation width restricts the number of effective speckle image samples available for model training, thus limiting the performance optimization space of deep learning models. Furthermore, it effectively suppresses the intermodal coupling effect caused by multimode fiber bending, thereby improving the model's speckle image feature extraction capability and the accuracy of identifying the corresponding incident light wavelength.

[0046] In some embodiments, the spectrometer further includes an optical fiber coupler for coupling the circularly polarized light output from the first optical fiber collimator to the multimode optical fiber. By placing the optical fiber coupler between the first collimator and the multimode optical fiber, a low-loss power connection can be achieved between the first collimator and the multimode optical fiber. In other embodiments, the spectrometer may not use an optical fiber coupler, but instead employs mechanical beam alignment to align and transmit the circularly polarized light output from the first optical fiber collimator to the multimode optical fiber.

[0047] In some embodiments, the multimode fiber includes a ring-core multimode fiber. Using a ring-core multimode fiber as a dispersive element not only enables the transmission of incident vortex beams of different orders, but also results in weak coupling between modes within the fiber, leading to a speckle pattern that is more stable than that of a typical graded-index multimode fiber. By receiving different speckle images at the tail end of the ring-core fiber, this method allows for the reception of more speckle images at a fixed wavelength, thereby enhancing the training sample data in the speckle image dataset.

[0048] This invention also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the speckle recognition and deep learning-based spectral measurement method described in any of the foregoing embodiments. The computer-readable storage medium can be a tangible storage medium, such as random access memory (RAM), main memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, floppy disks, hard disks, removable storage disks, CD-ROMs, or any other form of storage medium known in the art.

[0049] This invention also provides a computer program product, including computer instructions, which, when executed by a processor, implement the steps of the speckle recognition and deep learning-based spectral measurement method described in any of the foregoing embodiments.

[0050] Those skilled in the art will understand that the exemplary components, systems, and methods described in conjunction with the embodiments disclosed herein can be implemented in hardware, software, or a combination of both. Whether implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this invention. When implemented in hardware, it can be, for example, electronic circuits, application-specific integrated circuits (ASICs), appropriate firmware, plug-ins, function cards, etc. When implemented in software, the elements of this invention are programs or code segments used to perform the desired tasks. The programs or code segments can be stored in a machine-readable medium or transmitted over a transmission medium or communication link via data signals carried in a carrier wave.

[0051] It should be clarified that the present invention is not limited to the specific configurations and processes described above and shown in the figures. For the sake of brevity, detailed descriptions of known methods are omitted here. In the above embodiments, several specific steps are described and shown as examples. However, the method process of the present invention is not limited to the specific steps described and shown. Those skilled in the art can make various changes, modifications, and additions, or change the order of steps, after understanding the spirit of the present invention.

[0052] In this invention, features described and / or illustrated for one embodiment may be used in the same or similar manner in one or more other embodiments, and / or combined with or in place of features of other embodiments.

[0053] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, various modifications and variations of the embodiments of the present invention are possible. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A spectral measurement method based on speckle recognition and deep learning, characterized in that, The method includes: The speckle image time series data is input into a deep learning model to output the wavelength time series variation law of the incident light under test corresponding to the speckle image time series data. The speckle image time series data is generated by the incident light under test under arbitrary incident vortex order conditions. The deep learning model includes a non-convolutional neural network model. Based on the speckle image time-series data, the temporal variation law of the light intensity of the incident light to be measured corresponding to the speckle image time-series data is determined; based on the wavelength time-series variation law and the light intensity time-series variation law, the spectrum of the incident light to be measured is reconstructed.

2. The method according to claim 1, characterized in that, The deep learning model is pre-trained through the following steps: A pre-defined deep learning model is trained based on multiple speckle image data with their own incident light wavelength labels, so that the trained deep learning model can output the wavelength of the incident light corresponding to the speckle image data. The multiple speckle image data are generated by the excitation of multiple wavelengths of incident light under multiple incident vortex light order conditions. Each speckle image data is the average value of multiple speckle image data generated by the excitation of multiple fiber deformation conditions corresponding to each incident vortex light order.

3. The method according to claim 2, characterized in that, Before the step of training a pre-defined deep learning model based on multiple speckle image data each labeled with its own incident light wavelength, the training process further includes: Multiple speckle image data are preprocessed to obtain multiple preprocessed speckle image data, wherein the preprocessing includes removing the black parts in the speckle image data.

4. The method according to claim 2, characterized in that, The incident light of the multiple wavelengths is obtained through multiple samplings, and the wavelength sampling interval is determined by estimating the minimum spectral resolution of the reconstructed spectrometer.

5. The method according to any one of claims 1 to 4, characterized in that, The non-convolutional neural network model includes a visual model based on the Transformer architecture, which includes the Swin Transformer model.

6. A computational reconstruction spectrometer, characterized in that, The spectrometer includes: a tunable laser, a first polarizer, a spatial light modulator, a second polarizer, a first fiber collimator, a multimode fiber, a second fiber collimator, a near-infrared charge-coupled device camera, and a computer device. The tunable laser is used to provide incident light; The first polarizer is used to modulate the incident light into polarized light with a specific polarization state, and output circularly polarized light; The spatial light modulator is used to load a controllable order spiral wavefront onto the circularly polarized light, and output circularly polarized light under arbitrary incident vortex light order conditions. The second polarizer is used to modulate the circularly polarized light output by the spatial light modulator into polarized light with a specific polarization state, and output circularly polarized light; The first fiber collimator is used to transmit the circularly polarized light output from the second polarizer to the multimode fiber; The multimode fiber is used to generate speckle at the output end of the second fiber collimator based on the circularly polarized light output from the first fiber collimator using mode interference characteristics. The second fiber collimator is used to convert the divergent light transmitted by the multimode fiber into collimated light. The near-infrared charge-coupled device camera is used to receive the speckle and record and output speckle image data; The computer device includes a processor, a memory, and computer instructions stored in the memory. The processor is used to execute the computer instructions, and when the computer instructions are executed, the spectrometer implements the steps of the speckle recognition and deep learning-based spectral measurement method as described in any one of claims 1 to 5.

7. The computational reconstruction spectrometer according to claim 6, characterized in that, The spectrometer further includes an optical fiber coupler and a manual two-dimensional translation stage. The optical fiber coupler is used to couple the circularly polarized light output from the first optical fiber collimator to the multimode optical fiber. The manual two-dimensional translation stage is used to move the multimode optical fiber parallel to the x-axis or y-axis according to multiple preset optical fiber displacement points, so that the multimode optical fiber undergoes different degrees of deformation and generates speckle under multiple optical fiber deformation conditions, so that the near-infrared charge-coupled device camera can record and output speckle image data under multiple optical fiber deformation conditions.

8. The computational reconstruction spectrometer according to claim 6, characterized in that, The multimode fiber includes a ring-core multimode fiber.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the steps of the speckle recognition and deep learning-based spectral measurement method as described in any one of claims 1 to 5.

10. A computer program product comprising computer instructions, characterized in that, When executed by a processor, the computer instructions implement the steps of the speckle recognition and deep learning-based spectral measurement method as described in any one of claims 1 to 5.

Images