Infrared spectrum imaging module, model training method, imaging method and application

Abstract

The invention provides an infrared spectrum imaging module, a model training method, an imaging method and application. The imaging module comprises a main body, a main lens, a relay lens and a metasurface encoding device; the main body is provided with a light path channel, and the main lens, the metasurface encoder and the relay lens are sequentially arranged in the light path channel in the light propagation direction; the metasurface coding device is located on the focal plane of the main lens and is provided with a plurality of metasurface coding units which are arranged in an array on the focal plane; a connecting structure is further arranged at the tail end of the light path channel and on the main body and used for enabling the main body to be fixedly connected with the infrared imaging device. The optical field encoder is creatively and independently packaged in the module instead of being integrated on an infrared imaging chip, expensive chip assets are protected, flexible adaptation and plug and play are achieved, the optical development difficulty and hardware cost of a spectral imaging system are greatly reduced, and rapid integration and popularization are facilitated.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of photoelectric detection and imaging, and particularly relates to an infrared spectral imaging module, a model training method, an imaging method and application. BACKGROUND

[0002] With the rapid development of information technologies such as big data and artificial intelligence, the importance of new-generation intelligent photoelectric detection technology is increasingly highlighted. Advanced photoelectric detection and recognition systems play a key role and become indispensable "eyes" and "brains" in the process of dynamic perception and intelligent recognition. Under this background, spectral imaging technology has become a core technology in the field of intelligent photoelectric detection because it can simultaneously obtain spatial information of a target and spectral information representing material properties, significantly improving the dimension and magnitude of imaging data.

[0003] In the field of spectral imaging technology, traditional spectrometers are mainly divided into three categories: dispersive, filter and Fourier transform. However, these technical paths all have significant defects: (1) Dispersive spectrometer: relying on the spatial dispersion principle of gratings or prisms, it needs to obtain spectral information point by point through mechanical scanning, resulting in a large system volume and low acquisition efficiency.

[0004] (2) Filter spectrometer: using tunable filters to achieve wavelength selection, but it has the problems of low light flux and limited spectral resolution, and frequent switching of filters is required for multi-band detection, which seriously restricts real-time performance.

[0005] (3) Fourier transform spectrometer: based on interference principle, it can obtain high signal-to-noise ratio spectrum, but it needs to precisely move a mirror in the time dimension, which requires high mechanical stability and is difficult to realize miniaturization.

[0006] The common defects of these three traditional technologies are mainly in three aspects: first, the mechanical scanning mechanism in the spatial or time dimension leads to high system complexity and limited response speed; second, the bulky core optical elements limit the portability of applications; and third, the coupling design of precise optical components and complex mechanical structures significantly increases the manufacturing cost, making it difficult to meet the large-scale popularization demand.

[0007] To break through the traditional bottlenecks, computational spectral imaging technology has emerged. This technology uses micro-nano structures such as metasurfaces to encode light fields through a "physical encoding + digital decoding" collaborative mechanism, and then reconstructs the spectrum through algorithms, realizing the miniaturization of the system.

[0008] Currently, the "on-chip integration" technology based on metasurfaces is a research hotspot, which directly prepares microstructures on the surface of the detector pixels. However, in applications targeting high-performance "cryogenic infrared detectors", this on-chip integration scheme faces great technological and engineering challenges: (1)Encapsulation of the package structure: The refrigeration type infrared detector (such as MCT, InSb, T2SL) must be packaged in a vacuum Dewar assembly and coupled with a refrigerator to work in a low-temperature environment. Once the detector is vacuum encapsulated, its interior becomes a closed system, and subsequent spectral function upgrades or modifications cannot be performed.

[0009] (2) Extremely high process risk and cost: The refrigeration infrared focal plane chip is extremely expensive and fragile. Secondary micro-nano processing on the bonded chip can easily cause chip damage, resulting in expensive detector scrap and difficult to control the yield.

[0010] (3) Serious conflict between micro-nano processing technology and detector temperature tolerance: The preparation of the super surface often involves film plating process, photoetching baking or high-temperature annealing process, which usually requires a higher temperature environment. However, the refrigeration type infrared imaging chip (especially the tellurium cadmium mercury material) and the indium column used for interconnection are extremely sensitive to temperature and have poor heat resistance. Forcedly performing super surface preparation process involving high temperature on the finished chip can easily cause the performance degradation of the detector material, the diffusion of dopant, and even the failure of the readout circuit interconnection.

[0011] In order to avoid the risk of on-chip integration, using the secondary imaging architecture of "main lens-intermediate image plane coding-relay lens" becomes a feasible alternative idea. That is, using the main lens to focus the image on the external encoder, and then relaying the coded image to the inside of the detector through the lens.

[0012] However, there is no standardized mature module product integrating the functions of "super surface coding" and "relay imaging" in the current market, and there are the following technical and market gaps: (1) The lack of "plug and play" integrated module: Most existing infrared optical accessories are only single lenses. If users want to upgrade the spectrum of existing refrigeration infrared cameras, they often need to build a loose experimental optical path and assemble the super surface coding sheet and infrared lens scattered in front of the detector. This non-integrated structure is bulky, has poor mechanical stability, and is difficult to ensure precise optical axis alignment between the encoder and the lens, making it difficult to meet the needs of engineering applications.

[0013] (2) Insufficient adaptability of algorithms and hardware: Modern computational spectral imaging increasingly relies on deep learning algorithms to correct the blur in the light field transmission process and reconstruct the spectrum. The training and deployment of deep learning models require a highly stable and parameter-fixed optical hardware front end as support. The existing discrete component assembly scheme cannot maintain consistent optical parameters (such as distance and coaxiality), resulting in poor generalization ability of the algorithm model and the inability to stably correct optical blur through software.

[0014] In summary, the industry urgently needs an infrared spectral imaging module that is independent of the detector package and has a plug-and-play feature. SUMMARY

[0015] The main purpose of the present application is to provide an infrared spectral imaging module, a model training method, an imaging method and an application to overcome the shortcomings of the prior art.

[0016] To achieve the above invention purposes, the present application adopts the following technical solutions: In a first aspect, the present application provides a non-invasive plug-and-play infrared spectral imaging module, which includes a main body, a main lens, a relay lens and a super surface coding device; the main body has a light path channel, and the main lens, the super surface coding device and the relay lens are arranged in the light path channel in sequence along the light propagation direction; The super surface coding device is located on the focal plane of the main lens and has a plurality of super surface coding units arranged in an array on the focal plane; at the end of the light path channel, the main body is also provided with a connecting structure for fixed connection of the main body with an infrared imaging device.

[0017] In a second aspect, the present application further provides a training method of an infrared spectral imaging decoding model, which is applied to an infrared spectral imaging system composed of the above non-invasive plug-and-play infrared spectral imaging module and an infrared imaging device, and includes: Continuous infrared light of a known waveband is provided to irradiate the infrared spectral imaging system and a standard spectrometer respectively to obtain multi-channel spectral measurement values and spectral true values, and a physical prior is established based on the correspondence between the multi-channel spectral measurement values and the spectral true values; A training scene is provided, and the infrared spectral imaging system is used to collect measured coded images, and a hyperspectral imaging scanner is used to collect true value infrared spectra to establish an image data set; Based on the physical prior, a plurality of ideal coded images are simulated and generated by using an ideal physical model, a basic reconstruction network is pre-trained to obtain an ideal reconstruction network; The ideal reconstruction network is fine-tuned by using the image data set to obtain an infrared spectral imaging decoding model.

[0018] In a third aspect, the present application further provides an infrared spectral imaging method, which is applied to an infrared spectral imaging system composed of the above non-invasive plug-and-play infrared spectral imaging module and an infrared imaging device, and includes: An infrared spectral imaging decoding model is obtained by using the above training method; A target coded image of a target scene is collected by using the infrared spectral imaging system; Input the target coded image into the infrared spectral imaging decoding model to obtain an analysis spectrum of the target scene.

[0019] In a fourth aspect, the present application also provides application of the above-mentioned infrared spectral imaging method in the field of photoelectric detection.

[0020] The non-invasive plug-and-play infrared spectral imaging module provided by the present application has great significance for realizing low-cost spectral upgrading of a refrigeration-type infrared imaging system.

[0021] Compared with the prior art, the present application has at least the following beneficial effects: General architecture, reducing the optical development threshold: using modular design, the optical design of the relay lens has no special customization requirements, and can use specially designed lenses or directly integrate mature commercial infrared lens groups (such as infrared microscopes or standard infrared lenses) as relay components. This architecture greatly reduces the optical development difficulty and hardware cost of the spectral imaging system, facilitating rapid integration and promotion.

[0022] Non-invasive design, protecting valuable chip assets: creatively encapsulating the light field encoder in the module instead of integrating it on the infrared imaging chip. Users do not need to disassemble the expensive refrigeration-type infrared detector vacuum package, nor do they need to bear the high risk of chip micro-nanofabrication. They only need to connect the module provided by the present application like adding a "close-up adapter" or "magnifying lens" to make ordinary infrared cameras have spectral imaging functions.

[0023] Flexible adaptation, plug-and-play: the module has a mechanical adjustment margin, which can eliminate the flange distance error caused by different connection structures through simple mechanical focusing. Combined with the deep learning correction training method provided by the present application, even if the module uses ordinary lenses with slight aberrations, it can also be adaptively compensated through algorithms, realizing the perfect combination of low-cost hardware and high-performance imaging.

[0024] The above description is only a summary of the technical solutions of the present application. In order to enable those skilled in the art to more clearly understand the technical means of the present application and can be implemented according to the content of the specification, the following describes the preferred embodiments of the present application with reference to the detailed description of the drawings. DETAILED DESCRIPTION

[0025] In order to more clearly illustrate the technical solutions in the embodiments or the prior art, the following will briefly introduce the drawings needed to be used in the embodiment or prior art description. Obviously, the drawings in the following description are only some embodiments described in the present application, and those skilled in the art can obtain other drawings according to these drawings without creating any inventive labor.

[0026] Figure 1 is a schematic diagram of the interlayer structure of the metasurface coding device provided by an exemplary embodiment of the present application; Figure 2 is a schematic diagram of the microstructure of the metasurface coding device provided by an exemplary embodiment of the present application; Figure 3 is a schematic diagram of the architecture of the infrared spectral imaging decoding model provided by an exemplary embodiment of the present application; Figure 4a is a schematic diagram of a planar pattern structure of the metasurface coding device provided by an exemplary embodiment of the present application; Figure 4b is a schematic diagram of another planar pattern structure of the metasurface coding device provided by an exemplary embodiment of the present application; Figure 4c is a schematic diagram of still another planar pattern structure of the metasurface coding device provided by an exemplary embodiment of the present application; Figure 4d is a schematic diagram of yet another planar pattern structure of the metasurface coding device provided by an exemplary embodiment of the present application; Figure 5 is a schematic diagram of the structure design screening process of the metasurface coding unit of the metasurface coding device provided by an exemplary embodiment of the present application; Figure 6a is a simulation diagram of the correlation between the absolute correlation sum and the design height of the metasurface coding unit provided by an exemplary embodiment of the present application; Figure 6b is a simulation diagram of the correlation between the average transmittance and the design height of the metasurface coding unit provided by an exemplary embodiment of the present application; Figure 7a is a simulation diagram of the absolute cross-correlation numerical array of the metasurface coding unit provided by an exemplary embodiment of the present application when the design height is fixed at 2 μm; Figure 7b is a distribution diagram of the transmittance of each channel of the metasurface coding unit provided by an exemplary embodiment of the present application when the design height is fixed at 2 μm; Figure 8a is a simulation diagram of the correlation between the average correlation coefficient and the design channel number of the metasurface coding unit provided by an exemplary embodiment of the present application; Figure 8b is a simulation diagram of the correlation between the average transmittance and the design channel number of the metasurface coding unit provided by an exemplary embodiment of the present application; Figure 9 is a schematic diagram of the preparation process of the metasurface coding device provided by an exemplary embodiment of the present application; Figure 10is a structural schematic diagram of an infrared spectral imaging system provided by a typical embodiment of the present application; Figure 11 is a light path system schematic diagram of an infrared spectral imaging system provided by a typical embodiment of the present application; Figure 12a is an example diagram of fitting effects of a spectral reconstruction training set and a test set in a training method provided by a typical embodiment of the present application; Figure 12b is another example diagram of fitting effects of a spectral reconstruction training set and a test set in a training method provided by a typical embodiment of the present application; Figure 12c is another example diagram of fitting effects of a spectral reconstruction training set and a test set in a training method provided by a typical embodiment of the present application; Figure 12d is still another example diagram of fitting effects of a spectral reconstruction training set and a test set in a training method provided by a typical embodiment of the present application; Figure 13a is an example diagram of spectral reconstruction of a continuous light source with a half-peak width of 50 nm in a training method provided by a typical embodiment of the present application; Figure 13b is an example diagram of spectral reconstruction of a continuous light source with a half-peak width of 100 nm in a training method provided by a typical embodiment of the present application; Figure 13c is an example diagram of spectral reconstruction of a continuous light source with a half-peak width of 150 nm in a training method provided by a typical embodiment of the present application; Figure 13d is an example diagram of spectral reconstruction of a continuous light source with a half-peak width of 200 nm in a training method provided by a typical embodiment of the present application; Figure 14 is a contrast diagram of spectral imaging recovery effects in a single channel in a training method provided by a typical embodiment of the present application. DETAILED DESCRIPTION

[0027] Although the computational spectral imaging technology has covered various paths such as image plane coding and PSF (Point Spread Function) coding in the theoretical level, in the actual hardware engineering implementation, the existing solutions mainly have the following three forms, and none of them can meet the plug-and-play upgrade requirements of the cooled infrared detector: Special integrated machine: this kind of scheme integrates the light splitting elements (such as filter wheel, interferometer, digital micromirror light path) with the detector and imaging lens in a large machine box to form a special spectral camera.

[0028] The limitation of this kind of scheme is that: Closed and high cost: Users must purchase expensive whole machines, and cannot upgrade using existing ordinary infrared cameras.

[0029] Lack of flexibility: The spectral channels of the system are usually fixed at the factory, and cannot be replaced according to task requirements.

[0030] Bulkiness: Contains complex mechanical transmission or long optical path structure, and does not have the characteristics of lightweight modules.

[0031] On-chip integration: The micro-nano spectral structure is directly etched or coated on the surface of the detector focal plane chip.

[0032] Its limitations for modularization are: Invasive process: This is an internal modification of the detector manufacturing process, rather than an independent optical accessory. For packaged refrigeration infrared detectors, because they cannot be disassembled, this solution cannot be flexibly adjusted.

[0033] Non-independent product: The spectral structure is permanently bound to the chip, and users cannot replace the spectral encoder like a lens.

[0034] Discrete element building: This is a common "external" form in current laboratory research. Researchers usually use standard infrared lenses, independent metasurface coding sheets, and independent lens groups on an optical platform to build a loose experimental optical path through guide rails and supports.

[0035] Its limitations for modularization are: Non-standardized product: This form is only an "experimental device", not an "industrial product". There is no finished module on the market that encapsulates "metasurface encoder devices" and "relay lenses".

[0036] Engineering difficulties: The coaxiality and spacing distance between discrete elements are extremely sensitive. In the absence of integrated packaging, it is extremely difficult for users to achieve precise alignment of metasurfaces and relay lenses through manual adjustment, resulting in unstable imaging quality and the inability to resist vibration and impact, making it difficult to apply to actual engineering sites.

[0037] In summary, although the above existing computational spectral imaging technologies have made some progress in theoretical research, they still face insurmountable technical contradictions in practical engineering applications for refrigeration-type infrared imaging. Specifically, the existing technical solutions have the following significant defects: (1) Process compatibility and yield cost issues: The current on-chip integrated scheme based on metasurfaces requires micro-nano processing directly on the surface of the bonded refrigeration infrared focal plane chip. Due to the fragile texture and temperature sensitivity of tellurium cadmium mercury (MCT) and other infrared materials, the invasive etching and coating process is extremely easy to cause damage to the detector pixels, resulting in pixel damage and even chip scrap. This high-risk process at the "chip level" significantly increases the manufacturing cost and cannot guarantee the yield rate of mass production.

[0038] (2) System volume and integration issues (for image plane coding / mechanical light splitting): The system using a filter wheel or an interferometer introduces a bulky mechanical moving part, which not only increases the weight of the whole machine, but also reduces the integration of the system.

[0039] (3) Limited use scenarios and real-time performance (for time sequence acquisition / mechanical type): The filter wheel or tunable filter uses a "time-sharing scanning" mechanism, and the data of each spectral channel cannot be obtained at the same time. This non-snapshot acquisition method makes it only suitable for static scene observation. In the face of high-speed moving targets or transient combustion processes, serious spatial-spectral artifacts will be produced, resulting in failure of spectral reconstruction.

[0040] (4) Lack of task adaptability (for fixed design): Whether it is a metasurface chip integrated on-chip or an image plane coding system with a fixed mask, the spectral response characteristics are fixed after manufacturing. Once the detection task changes (such as changing the detection wavelength or adjusting the spectral resolution), users cannot simply replace the spectral functional unit like changing the lens, resulting in poor device versatility and low hardware reuse rate.

[0041] (5) Robustness issues (for loose external scheme): Existing external experimental devices are usually loosely built by discrete components, lacking integrated module design. The discrete structure is extremely sensitive to mechanical vibration and impact, and the optical axis is extremely easy to lose alignment, making the system less robust and difficult to work stably in harsh outdoor or industrial environments for a long time.

[0042] Therefore, in view of the deficiencies in the prior art, after long-term research and a large number of practices, the technical scheme of the present application is proposed. The technical scheme, its implementation process and principles will be further explained as follows.

[0043] In the following description, many specific details are set forth in order to provide a thorough understanding of the present application, however, the present application can also be implemented in other ways different from those described herein, therefore, the scope of protection of the present application is not limited by the specific embodiments disclosed below.

[0044] The present application aims at the core problems of high process risk, large system volume, and unstable external light path in the application of existing computing mid-infrared spectral imaging in refrigeration type detector, based on the combination of super surface technology, light field coding technology, algorithm decoding technology and module level integration technology, a non-invasive plug-and-play infrared spectral imaging module is designed, and the corresponding hardware architecture and supporting algorithm are proposed, the purpose and corresponding solution mainly include: (1) Solve the contradiction between high-risk process and low-cost upgrade, aiming to solve the yield risk and high cost problem caused by the traditional "on-chip integration" scheme to the expensive refrigeration infrared imaging chip. The solution proposed by the present application is to package the light field coding device (the super surface coding device provided by the present application belongs to one of the "light field coding devices") independently from the detector, and to select a material system compatible with mainstream optical processing and semiconductor process to independently prepare the super surface coding device, which not only greatly reduces the invasiveness risk to the detector body, but also significantly reduces the overall preparation cost and process implementation complexity of the spectral imaging system.

[0045] (2) Solve the stability and integration of external light path, aiming to solve the problem of large volume and instability of existing external discrete experimental light path. Compared with the traditional loose optical building scheme, the solution of the present application forms a miniaturized module by integrating the super surface coding device and the broadband infrared focusing lens. This "module level" integration realizes the miniaturization of the device and ensures the stability of the light field transmission.

[0046] (3) Improve the light energy utilization rate and coding independence, aiming to solve the light efficiency loss and signal aliasing caused by external coding. Through the carefully designed super surface micro-nano structure, the light utilization rate is greatly improved while realizing efficient spatial-spectral modulation of the light field; at the same time, through structure optimization design, the spectral coding correlation between super surface units is greatly reduced, ensuring that the detector can still receive distinctive coded information after module transmission.

[0047] (4) Realize high-resolution soft and hardware cooperative imaging, aiming to solve the problem of hardware blur and resolution limitation in mid-infrared spectral imaging. The present application deeply embeds the prior knowledge of the optical transfer function of the external module and the coding matrix into the algorithm model through the carefully designed supporting decoding algorithm. Using this "hardware coding + software correction" cooperative mechanism, the optical blur caused by the module can be effectively compensated, so as to realize high spectral resolution and high imaging resolution of mid-infrared spectral imaging.

[0048] Based on the above technical ideas, the embodiment of the present application first provides a non-invasive plug-and-play infrared spectral imaging module, which comprises a main body, a main lens, a relay lens and a metasurface coding device; the main body has a light path channel, the main lens, the metasurface coding device and the relay lens are sequentially arranged in the light path channel along the light propagation direction; the metasurface coding device is located on the focal plane of the main lens and has a plurality of metasurface coding units arranged in an array on the focal plane; at the end of the light path channel, a connecting structure is further arranged on the main body for fixed connection of the main body with an infrared imaging device.

[0049] In some embodiments, along the thickness direction of the metasurface coding device, the metasurface coding device sequentially comprises a substrate, a buffer layer and a metasurface micro-nano structure layer, the metasurface coding units are formed in the metasurface micro-nano structure layer, and the pattern satisfies four-fold rotational symmetry.

[0050] In some embodiments, the material of the substrate includes any one or a combination of two or more of silicon, germanium, sapphire, calcium fluoride and zinc selenide; the material of the buffer layer includes any one or a combination of two or more of silicon dioxide and aluminum oxide; the material of the metasurface micro-nano structure layer includes any one or a combination of two or more of alpha-Si, Ge, two-dimensional nanomaterials and quantum dots.

[0051] In some embodiments, the thickness of the buffer layer is 0.1-2 μm, and the thickness of the metasurface micro-nano structure layer is 1-5 μm.

[0052] In some embodiments, the inner wall of the light path channel is provided with an extinction surface.

[0053] The embodiment of the present application also provides a training method of an infrared spectral imaging decoding model, which is applied to an infrared spectral imaging system composed of the non-invasive plug-and-play infrared spectral imaging module provided in any of the above embodiments and an infrared imaging device, and comprises the following steps: Continuous infrared light of a known waveband is provided to irradiate the infrared spectral imaging system and a standard spectrometer respectively to obtain multi-channel spectral measurement values and spectral true values, and a physical prior is established based on the correspondence between the multi-channel spectral measurement values and the spectral true values; A training scene is provided, and a measured coding image is collected by using the infrared spectral imaging system, and a true value infrared spectrum is collected by using a hyperspectral imaging scanner to establish an image data set; Based on the physical prior, a plurality of ideal coding images are generated by using an ideal physical model simulation, a basic reconstruction network is pre-trained to obtain an ideal reconstruction network; The ideal reconstruction network is fine-tuned by using the image dataset to obtain an infrared spectral imaging decoding model.

[0054] In some embodiments, the training method can specifically include the following process: Based on the correspondence between the multi-channel spectral measurement value and the spectral true value, a spectral response matrix is formed; The spectral response matrix is embedded as a fixed weight layer in the front end of the basic reconstruction network; Numerical simulation is performed using a hyperspectral image database to generate the ideal encoded image, and the basic reconstruction network is trained; The backbone weight responsible for extracting the metasurface coding feature in the ideal reconstruction network is frozen, and the mapping layer parameter at the end of the ideal reconstruction network is updated by using the image dataset through a back propagation algorithm to obtain the infrared spectral imaging decoding model.

[0055] In some embodiments, the basic reconstruction network includes any one of a fully connected neural network, a U-Net neural network, and a Transformer neural network. Of course, the basic reconstruction network applicable to the present application is not limited to the above-mentioned several networks, and basically, a neural network containing an input head, an encoder, a decoder, and an output head (a discrimination head) can be used for the spectral reconstruction work of the present application.

[0056] Based on the above technical solution, the exemplary embodiments provided by the present application are divided into three parts: (1) super surface-based light field encoder device design, (2) preparation process of the light field encoder device and assembly with an infrared lens to form an infrared spectral imaging module, and (3) calibration test and supporting algorithm design. The technical solution of the present application will be specifically introduced from the three aspects below.

[0057] (1) Super surface-based light field encoder device design: The light field encoder device of the present application is composed of a substrate material corresponding to the mid-infrared waveband (selected from infrared optical materials with high transmittance in the mid-infrared waveband, including but not limited to silicon (Si), germanium (Ge), sapphire (Sapphire), calcium fluoride (CaF2), or zinc selenide (ZnSe), etc.), a buffer layer, and an alpha-Si metasurface micro-nano structure layer (the buffer layer can be any material with a buffering effect, such as silicon dioxide, aluminum oxide, etc., and the material of the metasurface micro-nano structure layer can also be any material with high transmittance efficiency in the mid-infrared wavelength, such as alpha-Si, Ge, two-dimensional materials, quantum dot materials, etc.). The light field encoder device is a super surface micro-nano structure layer (note that the structure layer here can be a multi-material stack with a corresponding super surface structure, and the thickness can be 1-5 μm) stacked on a buffer layer (the thickness can be 0.1-2 μm), and the schematic diagram of a single device is as follows: Figure 1As shown, this simple stacking combined with a carefully designed micro / nano structure (the design principle is mainly the near-field coupling of the internal structure of the structural unit) is easy to fabricate while simultaneously encoding the incident infrared light into a light field, thus forming a highly efficient and highly differentiated light field encoding effect.

[0058] For the design of metasurface micro / nano structure layers, a large number of structural units with different scanning parameters are used to calculate the transmittance and diffraction efficiency of various metasurface structures (circular, rectangular, cross-shaped, etc., the structure only needs to satisfy fourfold rotational symmetry, or C4 structure simulation can be performed by deep learning random QR code encoding) and various structural parameters (length, width, height). The transmission response with low similarity, high efficiency and high diffraction efficiency is selected by greedy algorithm to obtain the corresponding metasurface structure parameter results (the simulation process uses simulation software such as COMSOL).

[0059] Of course, the choice of substrate, buffer layer and metasurface micro / nano structure materials can be changed according to the required application band. For example, when a material that is transparent in the visible light band is selected, it can be used as a light field coding device in the visible light band, and it is expected to realize cross-band spectral coding devices and even cross-band spectral imaging applications.

[0060] (2) Fabrication process of the optical field encoding device and its assembly with an infrared lens to form an infrared spectral imaging module: First, the substrate material corresponding to the mid-infrared band (selecting infrared optical materials with high transmittance in the mid-infrared band, including but not limited to silicon (Si), germanium (Ge), sapphire, calcium fluoride (CaF2) or zinc selenide (ZnSe)) is simply cleaned. Then, a 0.1 μm thick SiO2 layer is grown on the substrate of the device as a buffer layer (the material is selected from dielectric materials such as silicon dioxide (SiO2) and aluminum oxide (Al2O3), mainly to be compatible with the structural layer material and the band, and the thickness can be selected from 0.1 μm to 2 μm). Then, a 1 μm thick α-Si material is deposited on the buffer layer as a metasurface micro-nano structure layer. Then, positive photoresist AZ1500 is sprayed, and the device is patterned by photolithography and development technology (any photolithography machine model can be used, and different resolution photolithography machines can be selected according to the actual application requirements for the size of the unit device). Next, Ti / Au / Cr was deposited using an electron beam evaporation deposition system, with Cr serving as the hard mask material for etching α-Si. Using an NLD etching system with Cr as the mask, the α-Si material outside the patterned area was etched away. After etching, the sample was immersed in acetone solution for 12 hours. The surface was gently blown through the solution using a dropper, and ultrasonicated for 30 seconds × 4 times to completely remove the photoresist, obtaining the desired metasurface micro / nano structure. Finally, the sample was rinsed several times with isopropanol and deionized water, and dried with nitrogen gas to complete the preparation. The specific structure after preparation is shown below. Figure 2 As shown.

[0061] After completing the fabrication of the metasurface coding device, it is integrated with a commercial lens into a single module. The specific steps are as follows: Step 1: Centering and packaging of metasurface coding devices.

[0062] The metasurface encoded device is mounted into a high-precision metal mirror mount. A micro-alignment system is used to assist in fine-tuning the planar position of the device, ensuring that the center of the micro / nano structure array coincides with the mechanical axis of the mirror mount (eccentricity error is controlled within the micrometer level), and it is fixed by a pressure ring or optical UV adhesive.

[0063] Step 2: Modular integration of relay lens group.

[0064] Inside the main lens barrel of the module, a selected commercial infrared lens group (or a standard broadband achromatic lens) is installed according to the preset object-image conjugate distance. By adjusting the length of the internal spacer or using a threaded focusing mechanism, the axial spacing between the lens group and the metasurface encoding device is precisely set to ensure that the metasurface plane is located at the object-side focal plane of the relay lens.

[0065] Step 3: Inner wall extinction and stray light suppression.

[0066] To eliminate ghosting caused by reflections of non-imaging beams on the inner walls of the module, all non-light-transmitting surfaces inside the module (such as the inner walls of the lens barrel, the surface of the spacer ring, and the surface of the pressure ring) undergo matting treatment. Treatment methods include coating with high-absorption infrared matting paint, attaching matting felt, or sandblasting and oxidation blackening treatment to maximize module performance.

[0067] Step 4: Standardized interface assembly and coaxiality testing.

[0068] Standardized mechanical interfaces are installed at both ends of the module (the front end is for the main lens, and the rear end is for the detector). After assembly, the imaging quality of the module is observed using a collimator to check for any obvious assembly misalignment or tilt, ensuring that the optical axis of each optical element is aligned with the mechanical axis.

[0069] (3) Calibration testing and supporting algorithm design: To eliminate the discrepancy between simulation and reality, a real-world dataset containing accurate optical aberration information needs to be constructed for subsequent fine-tuning and correction of the network. Spectral response calibration set (1D point data): For calibration of the spectral detection dimension, mid-infrared continuous light sources (such as blackbody radiation sources or tunable lasers) with different frequency response characteristics are used, and incident on the module of this invention and the standard spectrometer system respectively. The readings of the standard spectrometer are used as the "true spectral value", and the response data of each channel read by the detector of this module are used as the "measured value" to construct a basic spectral response mapping dataset.

[0070] Spectral Image Calibration Set (3D Cubic Data): A high-precision data acquisition system is built for calibration of the spectral imaging dimension. First, a series of static scenes are scanned using a standard hyperspectral imaging scanner (or a standard narrowband filter wheel is added in front of the detector) to acquire high spatial resolution spectral image data cubes, which serve as the "ground values" for network training. Subsequently, keeping the scene unchanged, the plug-and-play module of this invention is loaded onto an infrared detector to take a single snapshot of the same scene, generating a corresponding "fuzzy coded image" carrying light field encoding information and lens blur information.

[0071] The reconstruction algorithm of this invention (e.g.) Figure 3 (As shown) It is not limited to a specific network architecture (it can use fully connected neural networks, U-Net, or Transformer, etc.), and its core innovation lies in the "post-correction" training strategy: Phase 1: Offline Pre-training Based on Optical Field Encoding Priors. An ideal physical model is constructed based on the theoretical transmittance function (design value) of the metasurface-based optical field encoding device. Numerical simulations are performed using a massive hyperspectral image database to generate coded images under ideal conditions, which are then used to train and reconstruct the network. In this phase, the network utilizes optical field encoding as a strong physical prior, learning the general laws for deconstructing the spectrum from encoded aliasing information.

[0072] Phase Two: In-situ Post-Correction Based on Measured Data. Addressing aberrations and assembly errors introduced by the relay lens in the actual module, a small amount of "measured ground truth - blurred coded image" data is used to perform transfer learning or parameter fine-tuning on the pre-trained base network. Specifically, the backbone weights responsible for extracting metasurface coding features are frozen, and the mapping layer parameters at the network's end are updated using the measured data via backpropagation. This step is equivalent to performing "software adaptive correction" at the algorithm level, effectively compensating for the optical blurring errors in the hardware during computation.

[0073] After completing network fine-tuning, a comprehensive performance evaluation of the prototype was conducted: Spectral detection accuracy verification: After the module is calibrated with continuous light source incident, the spectral reconstruction results are obtained and compared with the test curve of a standard spectrometer (the spectral angle plotting SAM or root mean square error RMSE is calculated) to quantitatively evaluate its single-point spectral detection accuracy.

[0074] Spectral imaging quality verification: Using the established acquisition system, an unknown scene is imaged onto an infrared imaging chip equipped with this invention. The acquired coded data is input into a post-corrected algorithm model for reconstruction. The reconstructed hyperspectral data cube is compared with reference data acquired by a standard instrument in multiple dimensions (such as spatial resolution and spectral fidelity) to verify the invention's ability to achieve high-resolution spectral imaging in a non-invasive architecture. An exemplary spectral imaging reconstruction algorithm architecture is as follows: Figure 3 As shown (image rotated 90° to the left).

[0075] This invention also provides an infrared spectral imaging method, which is applied to an infrared spectral imaging system composed of a non-invasive plug-and-play infrared spectral imaging module and an infrared imaging device provided in any of the above embodiments, and includes the following steps: The infrared spectral imaging decoding model is obtained by using the training method provided in any of the above embodiments; The infrared spectral imaging system is used to acquire target-coded images of the target scene; The target coded image is input into the infrared spectral imaging decoding model to obtain the analytical spectrum of the target scene.

[0076] The embodiments of the present invention also provide the application of the infrared spectral imaging method provided in any of the above embodiments in the field of photoelectric detection.

[0077] The technical solution of the present invention will be further described in detail below through several embodiments and in conjunction with the accompanying drawings. However, the selected embodiments are only for illustrating the present invention and do not limit the scope of the present invention.

[0078] Example 1 This invention discloses a non-invasive, plug-and-play infrared spectral imaging module design, corresponding fabrication process, and supporting algorithms. By designing and fabricating a metasurface-based optical field encoding device on an independent optical substrate, and encapsulating it with a broadband achromatic relay lens into an integrated module, a non-destructive spectral upgrade of commercially available cooled infrared detectors is achieved. Through extensive scanning of micro / nano structures satisfying four-fold rotational symmetry (C4 symmetry), and combining a greedy algorithm to calculate the spectral response correlation between different structural units, the combination of structural units with the lowest correlation and highest light energy utilization is selected as the mask array for the optical field encoding device. Finally, the encoded data acquired by this module is used as input to a deep learning network, and high-resolution spectral reconstruction and imaging are achieved by incorporating physical priors. The following will describe specific embodiments of the design and fabrication, spectral reconstruction, and spectral imaging: (1) Design of optical field encoding device: The design principle of the metasurface micro / nano structure unit in this invention mainly utilizes the change in broadband response caused by the near-field coupling of the internal structure of the micro / nano unit. This can be achieved by performing simple material stacking on the detector pixel and then selectively etching the specific functional layer. Its side view is shown in the figure. Figure 1 , Figure 2 As shown, top views of various simplified structural designs are available. Figures 4a-4d As shown in the figure. (In the figure, P is the period, a is the duty cycle, and there is also a parameter not shown in the figure: height h). Here, the parameters are set as follows: period P is 1μm-5μm, and duty cycle a is 0.1-0.9. (The parameter selection here is mainly to match the wavelength, because the essence of metasurfaces is to control the light field through micro-nano structure units on the order of half wavelength. Considering the processing difficulty and the actual application wavelength, the above parameters are used for scanning.)

[0079] The micro / nano structural units were simulated using COMSOL multiphysics simulation software. The incident light source was set as a superposition of TE and TM polarized light. Periodic boundary conditions were added on both sides of the structural units to simulate the electromagnetic properties of the micro / nano structural units in an infinitely large periodic structure. After scanning the structural parameters, a large number of broadband response results were obtained. Then, a greedy algorithm was used to select a number of micro / nano structural units with low correlation, which were then used as optical field encoding devices and detectors for integrated integration.

[0080] Greedy algorithm-based selection strategies such as Figure 5 As shown.

[0081] In the simulation structural parameters, six different heights were selected: 0.5 μm, 1 μm, 1.5 μm, 2 μm, 2.5 μm, and 3 μm. Extensive simulations were performed on metasurfaces at different heights, with other parameters kept constant. For each of the six heights, 12,000 structures were simulated, resulting in a total of 72,000 data points. To select a suitable observation matrix, where the correlation between any two vectors (i.e., the spectral response between two structures) should be as low as possible (below 0.1), a simplified calculation process was adopted due to the large amount of data. The mean correlation and the mean transmittance of the entire spectrum were used for evaluation.

[0082] Data was extracted using Matlab, summarized by height, and then 100 structures were selected. The selection algorithm was designed based on a greedy strategy. By calculating the sum of their cross-correlation coefficients and the broadband average transmittance, a total of 100 structures in ten groups were selected for each of six different heights. The results are as follows: Figure 6a and Figure 6bAs shown. Considering factors such as processing difficulty, transmission efficiency, and sufficiently low correlation, a column height of 2 micrometers is most suitable. At this height, the average cross-correlation coefficient of the metasurface is below 0.2, and the average efficiency is above 80%. Figure 7a and Figure 7b As shown.

[0083] After determining the height of the processing unit, the simulation unit structure was further enriched based on a height of 2μm. Subsequently, a greedy strategy was used to screen for different numbers of structure groups. Structure group screening was conducted with 9, 16, 25, 36, 49, 64, 81, and 100 sub-units. For each number of structure groups, the optimal 10 groups were selected, and the mean correlation coefficient and mean transmittance of different numbers of structure groups were calculated. Figure 8a and Figure 8b As shown.

[0084] Considering factors such as the mean correlation coefficient, the mean transmittance, and the training effect of deep learning spectral reconstruction, 36 structure groups are used as broadband filtering matrices for transmittance metasurfaces.

[0085] (2) Fabrication process of the optical field encoding device: In a preferred embodiment of the present invention, the traditional high-risk on-chip integration scheme is abandoned, and instead a technical route of "independent device fabrication + modular packaging" is adopted. The metasurface is fabricated on an independent optical substrate compatible with mainstream semiconductor processes, which not only avoids damage to expensive detectors but also reduces the complexity of the process implementation. The fabrication mainly uses plasma-enhanced chemical vapor deposition (PECVD), electron beam lithography, magnetic neutral loop discharge plasma etching (NLD), and other instruments. The main process steps are: substrate preparation, PECVD deposition of SiO2 thin film, PECVD deposition of α-Si thin film, spin coating, electron beam lithography, electron beam evaporation of the top Cr layer, lift-off, NLD etching, and Cr removal. See the detailed process flow diagram below. Figure 9 As shown (image rotated 90° to the left).

[0086] 1) Substrate Preparation: A double-polished GaSb wafer transparent to the mid-infrared band is selected as the independent substrate, instead of using the original detector wafer directly. A simple substrate cleaning step is sufficient. The typical substrate cleaning steps are as follows: First, take a glass beaker and a PTFE basket of appropriate capacity, rinse them repeatedly with deionized water, and blow dry the inner wall of the beaker and the basket with a nitrogen gun to remove excess moisture; then take out the sample and put it into the basket, add acetone solution to the beaker, immerse the basket and the sample together in the solution, and then put it into an ultrasonic cleaner for 5 minutes. Repeat this operation twice; next, change to isopropanol solution and clean the substrate for 5 minutes, repeating twice, with other operations the same as the previous step; finally, rinse the substrate repeatedly with deionized water, blow dry with nitrogen, and put it into the sample box.

[0087] 2) PECVD deposition of SiO2 thin film as buffer layer: SiO2 thin film with a thickness of 0.24μm (the thickness can be selected from 0.1μm to 2μm) is prepared using a PECVD system. The growth rate is 120nm / min, the growth time is 2min, and the working temperature is 100℃.

[0088] 3) PECVD deposition of α-Si thin film as metasurface micro / nano structure layer (film layer before micro / nano processing): 3μm thick α-Si thin film is prepared using PECVD system (the thickness can be selected from 1μm to 3μm, mainly according to the parameters of the designed structure), with a growth rate of 300nm / min, a growth time of 10min, and an operating temperature of 250℃.

[0089] 4) Spin coating: Use A4 photoresist specifically for electron beam lithography. First, run the spindle at 600 rpm for 6 seconds, then run it at 4000 rpm for 30 seconds to homogenize the coating. Then bake at 180℃ for 10 minutes.

[0090] 5) Electron beam lithography: Any type of electron beam lithography machine can be used, and lithography can be performed using specific dosage process parameters at a speed of 1200 μm / min.

[0091] 6) Development: After exposure, develop using a mixed solution for 100 seconds and fix for 30 seconds.

[0092] 7) Apply glue: Use a glue applicator, 200W power, for one minute.

[0093] 8) Electron beam evaporation of top layer Cr: Cr is deposited using an electron beam evaporation coating machine with a thickness of 50nm.

[0094] 9) Metal stripping: After coating, immerse the sample in NMP solution for 12 hours. Then, place the sample in an ultrasonic cleaner and sonicate for 15 minutes, followed by a water bath at 85°C for 15 minutes. After the water bath, place the sample in an ultrasonic cleaner again and sonicate for 15 minutes.

[0095] 10) NLD Etching: Next, the structure of the optical field encoding device is fabricated by etching downwards, using Cr as a mask. Here, only an NLD etching machine is needed to etch away the excess α-Si in one go, and the etching time is set to 300s. Even if a little bit of the bottom SiO2 is etched away, the impact on the final result is very small and can be ignored.

[0096] 11) Removal of residual Cr: Immerse the prepared sample in Cr etching solution for 100 seconds to remove the residual Cr mask on the surface of the structure.

[0097] Assembly process of infrared spectral imaging module: After the individual components are fabricated, they need to be packaged with the relay lens into an integrated module. The assembly process focuses on the precise alignment of the optical axis and mechanical stability, and the specific steps are as follows: Step 1: Centering and packaging of metasurface coding devices.

[0098] Device packaging: The prepared metasurface-based optical field encoding device is installed in a high-precision metal lens holder, and a micro-alignment system is used to ensure that the geometric center of the micro-nano structure array coincides with the optical axis.

[0099] Front focal plane matching: To ensure that the imaging plane of the front-end infrared main lens (in this embodiment, a commercial broadband achromatic infrared lens with f=12.7mm and NA=0.67 is used) falls precisely on the metasurface coding device, its relative position to the metasurface needs to be adjusted according to the nominal focal length of the main lens. By adjusting the length of the module's front-end interface or the thickness of the internal gasket, the optical distance from the front-end interface plane to the metasurface micro / nano structure layer is strictly matched with the back focal length of the main lens, thereby ensuring that after external light is converged by the main lens, the focal point is accurately located on the metasurface coding plane.

[0100] Step 2: Modular integration of relay lens group.

[0101] Inside the main lens barrel of the module, behind the metasurface coding device, a broadband achromatic relay lens group is installed.

[0102] To construct a high-quality symmetrical optical relay architecture, this embodiment uses a commercially available infrared lens with parameters identical to the main lens (f=12.7mm, NA=0.67). This symmetrical design helps to minimize distortion and aberrations in the system.

[0103] Step 3: Conjugate distance adjustment and detector adaptation (these distances and lens selection can be changed according to actual needs).

[0104] Based on the actual packaging structure of the cooled infrared detector, the distance from the main plane of the relay lens to the photosensitive surface of the detector is set to 25.4 mm, i.e., the image distance v.

[0105] According to the Gaussian imaging formula, with a focal length f = 12.7 mm and an image distance v = 25.4 mm, the required object distance u is approximately 25.4 mm.

[0106]

[0107] By rotating the threaded focusing mechanism inside the module or replacing spacers of different lengths, the distance between the metasurface encoding device and the relay lens is adjusted to match the calculated object distance. This step ensures that the intermediate encoded image located on the metasurface can be clearly re-imaged onto the detector array at a distance of 25.4 mm from the lens.

[0108] Step 4: Inner wall matte finish and overall machine coaxiality test All non-optical light-transmitting surfaces inside the module (such as the inner wall of the lens barrel and the surface of the spacer) are sandblasted, oxidized and blackened or coated with infrared matte paint to suppress stray radiation.

[0109] After assembly, the front-end main lens and the back-end detector are connected, and the whole-machine imaging test is carried out using a collimator to confirm that the sharpness of the center and edge of the field of view is consistent, thus completing the module manufacturing.

[0110] The optical path diagram of the infrared spectral imaging system built based on this module is shown below. Figure 10 As shown.

[0111] (3) Spectral Reconstruction Implementation: To eliminate hardware errors and achieve high-precision reconstruction, this embodiment adopts an algorithm strategy of "physical prior driving + experimental fine-tuning". For example... Figure 11 As shown, a mid-wave infrared monochromatic light source (1.8–5.4 μm) is used to illuminate the module, and the detector directly reads the response of each channel to obtain the measured spectral transmission matrix (physical prior). A paired dataset is constructed by illuminating a standard spectrometer (to obtain the true value) and the module of this invention (to obtain the measured value) with a continuous light source. The mid-infrared spectral imaging data is first acquired using the pre-built acquisition system (requiring the use of a mid-infrared detector of the same source for filtering and acquisition of spectral imaging data in the corresponding band).

[0112] In this embodiment, a deep neural network is selected as the basic architecture for training the spectral reconstruction network, but it is not limited to this. The final test results of the spectral reconstruction network are as follows: Figures 12a-12d and Figures 13a-13d As shown.

[0113] For spectral imaging, U-Net, an attention mechanism, and residual connections are used as the basic model. The calibrated spectral response matrix of the optical field encoding device is embedded as a fixed weight layer in the network front end, giving the network a clear physical meaning of spectral dealiasing. The network is first pre-trained using a large amount of simulation data, and then transferred to the network using a small amount of measured data to adaptively compensate for PSF blurring and assembly errors introduced by the relay lens. The network's performance on the test set after training is shown below. Figure 14As shown in the figure, Peak Signal-to-Noise Ratio (PSNR) is a metric for measuring image quality based on the mean squared error (MSE) between corresponding pixels. It represents the ratio between the maximum possible power of the signal and the power of the noise that affects its representation accuracy. Structural Similarity Index Measure (SSIM) is a metric for measuring the perceptual similarity between two images. Unlike PSNR, which only focuses on the absolute error of pixel values, SSIM simulates the human visual system (HVS) and compares images from three dimensions: luminance, contrast, and structure. Spectral Angle Mapper Loss (SAM) primarily measures the fidelity of the spectral shape and is relatively insensitive to changes in illumination intensity; a lower SAM value indicates better performance.

[0114] As can be seen, using the spectral imaging system and supporting algorithm provided in this embodiment, the highest peak signal-to-noise ratio (PSNR) can reach 43.4 dB and the average structural similarity (maximum value 1) can reach 0.94 (maximum 0.99) when making predictions. The lowest PSNR can reach 33.9 dB and the average structural similarity is 0.89.

[0115] Based on the above embodiments, it can be clearly understood that: (1) Unlike traditional expensive and complex spectral imaging systems, this invention proposes a non-invasive, spliced ​​computational infrared spectral imaging module. This module creatively adopts a "split architecture" that integrates a metasurface-based optical field encoding device, a commercial relay lens, and a standard mechanical interface into one unit. This design does not require special customization of the internal lenses (commercially available infrared lenses can be used directly), and high-resolution spectral imaging can be achieved through simple physical cascading and algorithm decoding.

[0116] (2) The “independent modular” architecture brings remarkable engineering and cost advantages. This non-invasive design completely avoids the extremely high risks (such as high temperature damage and low yield) of directly performing micro-nano fabrication on expensive and fragile cooled infrared imaging chips. Compared with on-chip integration, this invention fabricates the core encoding device on an independent and inexpensive substrate, which not only greatly reduces the cost of trial and error, but also makes it possible to replace the encoder as easily as “changing a filter”, greatly reducing maintenance costs.

[0117] (3) Extremely low threshold for process implementation. The optical field encoding devices in the module are mass-produced on independent planar substrates (such as silicon wafers or gallium antimonide wafers) using mature semiconductor processes, while the relay imaging part is entirely produced using mature mechanical assembly processes. This "independent fabrication + mechanical assembly" manufacturing model does not require exploring complex heterogeneous integration or alignment processes and can be produced directly using the existing industrial supply chain, laying a solid foundation for rapid commercialization.

[0118] (4) Extremely high compatibility with commercial optical components. The integrated module of this invention no longer relies on complex and expensive custom optical designs, but can directly integrate standard commercial infrared lens groups (such as conventional infrared microscope objectives or fixed-focus lenses). The module can be adapted to lenses and detectors of different specifications through simple mechanical structure adjustments (such as washers and threads). This "standardized component integration" approach greatly reduces the hardware threshold and development cycle of spectral imaging systems.

[0119] (5) This invention no longer pursues the ultimate optical hardware performance, but instead utilizes a matching deep learning algorithm to accommodate the shortcomings of the hardware. Through a post-correction strategy from simulation to reality, the algorithm can adaptively compensate for aberrations and blurring caused by ordinary commercial lenses and simple assembly. This allows the module to obtain high-quality spectral images through the algorithm without the need for precise optical design, truly realizing "using software to reduce hardware costs".

[0120] (6) The module provided in the embodiments of the present invention is only equivalent in size to a standard lens adapter ring, which greatly reduces the physical size of the spectroscopic equipment. It can be directly screwed onto existing handheld infrared thermal imagers, security monitoring heads or drone pods, instantly giving ordinary infrared cameras spectral detection capabilities, solving the problem of traditional spectroscopic equipment being bulky and unable to operate on a mobile basis.

[0121] (7) Excellent scalability and versatility. Thanks to the standardized design of the internal space of the module, the present invention has excellent versatility. By replacing the metasurface-based optical field encoding device inside the module (changing the band) or replacing the commercial relay lens with a different focal length, the module architecture can be seamlessly extended to any band or adapted to any array detector, and has extremely high technical reuse value.

[0122] It should be understood that the above embodiments are merely illustrative of the technical concept and features of the present invention, and are intended to enable those skilled in the art to understand the content of the present invention and implement it accordingly. They should not be construed as limiting the scope of protection of the present invention. All equivalent changes or modifications made in accordance with the spirit and essence of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A non-invasive, plug-and-play infrared spectral imaging module, characterized in that, It includes a main body, a main lens, a relay lens, and a metasurface encoder; the main body has an optical path channel, and the main lens, the metasurface encoder, and the relay lens are sequentially arranged in the optical path channel along the direction of light propagation; The metasurface coding device is located on the focal plane of the main lens and has multiple metasurface coding units arranged in an array on the focal plane. At the end of the optical path channel, the main body is also provided with a connecting structure for fixing the main body to the infrared imaging device.

2. The non-invasive plug-and-play infrared spectral imaging module according to claim 1, characterized in that, Along the thickness direction of the metasurface coding device, the metasurface coding device sequentially includes a substrate, a buffer layer and a metasurface micro / nano structure layer, and the metasurface coding unit is formed in the metasurface micro / nano structure layer, and the pattern satisfies fourfold rotational symmetry.

3. The non-invasive plug-and-play infrared spectral imaging module according to claim 2, characterized in that, The substrate material includes any one or a combination of two or more of silicon, germanium, sapphire, calcium fluoride, and zinc selenide. The material of the buffer layer includes any one or a combination of two or more of silicon dioxide and aluminum oxide. The material of the metasurface micro / nano structure layer includes any one or a combination of two or more of α-Si, Ge, two-dimensional nanomaterials, and quantum dots.

4. The non-invasive plug-and-play infrared spectral imaging module according to claim 2, characterized in that, The thickness of the buffer layer is 0.1-2 μm, and the thickness of the metasurface micro / nano structure layer is 1-5 μm.

5. The non-invasive plug-and-play infrared spectral imaging module according to claim 1, characterized in that, The inner wall of the optical path channel is provided with an matting surface.

6. A training method for an infrared spectral imaging decoding model, characterized in that, The infrared spectral imaging decoding model is applied to an infrared spectral imaging system composed of a non-invasive plug-and-play infrared spectral imaging module and an infrared imaging device as described in any one of claims 1-5, including: Continuous infrared light of known wavelengths is provided to illuminate the infrared spectral imaging system and the standard spectrometer respectively, to obtain multi-channel spectral measurement values ​​and true spectral values, and a physical prior is established based on the correspondence between the multi-channel spectral measurement values ​​and the true spectral values; A training scenario is provided, and the infrared spectral imaging system is used to acquire measured coded images, while a hyperspectral imaging scanner is used to acquire true infrared spectra to establish an image dataset. Based on the aforementioned physical priors, multiple ideal encoded images are generated using an ideal physical model to pre-train the basic reconstruction network, thereby obtaining the ideal reconstruction network. The parameters of the ideal reconstruction network are fine-tuned using the image dataset to obtain the infrared spectral imaging decoding model.

7. The training method according to claim 6, characterized in that, Specifically, it includes: Based on the correspondence between multi-channel spectral measurements and true spectral values, a spectral response matrix is ​​formed; The spectral response matrix is ​​embedded as a fixed weight layer in the front end of the basic reconstruction network; Numerical simulations were performed using a hyperspectral image database to generate the ideal encoded image, and the basic reconstruction network was trained. The backbone weights responsible for extracting metasurface coding features in the ideal reconstruction network are frozen, and the mapping layer parameters at the end of the ideal reconstruction network are updated using the image dataset through a backpropagation algorithm to obtain the infrared spectral imaging decoding model.

8. The training method according to claim 6, characterized in that, The basic reconstruction network includes any one of the following: fully connected neural network, U-Net neural network, and Transformer neural network.

9. An infrared spectral imaging method, characterized in that, An infrared spectral imaging system comprising a non-invasive plug-and-play infrared spectral imaging module and an infrared imaging device as described in any one of claims 1-5, comprising: An infrared spectral imaging decoding model is obtained using the training method described in any one of claims 6-8; The infrared spectral imaging system is used to acquire target-coded images of the target scene; The target coded image is input into the infrared spectral imaging decoding model to obtain the analytical spectrum of the target scene.

10. The application of the infrared spectral imaging method according to claim 9 in the field of photoelectric detection.

Images