Novel on-chip coding calculation type hyperspectral imager

Abstract

The invention discloses a novel on-chip coding calculation type hyperspectral imager, and relates to the technical field of spectral imaging. The system specifically comprises an optical imaging assembly which is used for carrying out space focusing on light of a target scene to form a two-dimensional light field; the spectrum coding module is used for modulating the spectrum information of the two-dimensional light field through a periodic micro-nano structure array to generate a coded light signal; the photoelectric sensor is used for converting the coded optical signal into digital image data; the spectrum reconstruction module is used for decoding the digital image data based on a deep learning algorithm and generating a three-dimensional data cube containing space-spectrum information; wherein the periodic distribution parameter of the micro-nano structure array is matched with the point spread function characteristic of the optical imaging assembly. The invention aims to reduce the manufacturing cost of a spectrum encoder, improve the production efficiency of the spectrum encoder and solve the matching problem of the spectrum encoder and an imaging system.

Description

Technical Field

[0001] The present invention relates to the field of spectral imaging technology, and particularly to a novel on-chip coded computational hyperspectral imager. Background Art

[0002] Hyperspectral imaging technology can simultaneously obtain information of a target in multiple different bands, thus providing important data support for spectral analysis of substances. It is widely used in fields such as space remote sensing, target recognition, substance analysis, industrial inspection, and medical diagnosis. Compared with traditional imaging technologies, hyperspectral imaging can provide more spectral information, greatly improving the recognition accuracy and analysis ability of targets, and thus playing an irreplaceable role in scientific research and practical applications. The rapid development of this technology has promoted innovative progress in multiple disciplines such as physics, chemistry, and biology, and brought a huge impetus to the development of related industries.

[0003] However, traditional hyperspectral imagers usually rely on long optical path optical systems and large dispersive components, which makes the equipment large in volume, high in manufacturing cost, and complex in development, restricting its popularization in some practical applications. Especially in application scenarios with high requirements for portability, integration, and cost, such as satellite remote sensing, deep space exploration, industrial inspection, etc., the volume and cost problems of traditional hyperspectral imagers are particularly prominent. With the continuous growth of hyperspectral imaging requirements, how to overcome the disadvantages of large volume and high cost of traditional imaging instruments while ensuring hyperspectral imaging quality has become an important topic in the development of current hyperspectral imaging technology.

[0004] As an important development direction of modern spectral technology, lightweight and miniaturized hyperspectral imagers are gradually becoming a key technology in scientific research and engineering applications. By using a micro-nano structure array as a spectral encoding device to replace traditional dispersive spectroscopic elements, the volume of the spectrometer can be significantly reduced, thereby realizing the miniaturization, low cost, and high integration of the instrument. The micro-nano structure array generates a specific spectral response to the incident spectrum through different materials and geometric shapes, thereby realizing the encoding and decoding of spectral information. This principle has been verified in micro-nano photonics and modern computational spectroscopy technology, and laid a technical foundation for the lightweight and miniaturization of hyperspectral imagers.

[0005] In addition to innovations in hardware, the progress of software technology also provides support for the realization of lightweight and miniaturized hyperspectral imagers. With the help of computer technology, efficient, stable, and fast calculation methods can be designed to cooperate with new spectral encoding devices for accurate spectral information reconstruction. Combining the coordinated development of hardware and software can significantly improve the performance of hyperspectral imagers and further expand their application scope to various fields, especially in remote sensing, environmental monitoring, and health diagnosis.

[0006] However, although nanophotonic structures have shown great potential in the field of hyperspectral imaging, the manufacturing process of related spectral encoding devices is still relatively complex and the production cost is high. In addition, it is often impossible to achieve complete hyperspectral imaging functions by relying solely on encoding devices, and further technological breakthroughs and innovations are still needed.

[0007] Therefore, how to reduce the manufacturing cost of spectral coding devices, improve their production efficiency, and solve the matching problem between them and the imaging system has become a technical problem that needs to be solved urgently. Summary of the invention

[0008] The main purpose of the present invention is to provide a novel on-chip coding computational hyperspectral imager, aiming to reduce the manufacturing cost of spectral coding devices, improve their production efficiency, and solve the matching problem between them and the imaging system.

[0009] In order to achieve the above object, the present invention proposes a novel on-chip coding computational hyperspectral imager, comprising: An optical imaging component is used to spatially focus the light of the target scene to form a two-dimensional light field; A spectral encoding module, which modulates the spectral information of the two-dimensional light field through a periodic micro-nanostructure array to generate a coded light signal; a photoelectric sensor for converting the coded light signal into digital image data; A spectral reconstruction module, which decodes the digital image data based on a deep learning algorithm to generate a three-dimensional data cube containing spatial-spectral information; Wherein, the periodic distribution parameters of the micro-nano structure array match the point spread function characteristics of the optical imaging component.

[0010] In one embodiment of the present application, the spectral encoding module includes: Transparent glass sheet substrate; An ultraviolet light-cured polymer sheet is disposed on the transparent glass substrate, and a structure array consisting of a plurality of nanophotonic hole arrays is formed on the ultraviolet light-cured polymer sheet; and The metal layer is deposited on the side of the ultraviolet light-cured polymer sheet away from the transparent glass substrate.

[0011] In one embodiment of the present application, the structure array is a 16*16 rectangle, and the aperture sizes and / or hole arrangement periods and / or hole arrangement shapes in each nanophotonic hole array are different.

[0012] In one embodiment of the present application, the aperture size of each nanophotonic hole array falls within a first preset range, which is 100 nm to 850 nm.

[0013] In an embodiment of the present application, the pore arrangement period within each nano - photonic pore array falls within a second preset range, and the second preset range is from 300 nm to 900 nm.

[0014] In an embodiment of the present application, the arrangement shape of each nano - photonic pore array is square or hexagonal.

[0015] In an embodiment of the present application, the deep - learning architecture adopts the basic CNN architecture, uses the response dataset generated by the target radiation spectrum irradiated by a standard light source as the training input, uses the response dataset generated by the standard spectrometer receiving the same spectrum as the ground - truth input, and uses the mean - square error value of the spectral difference between the two to embed the loss function of the optical PSF knowledge to train the spectral restoration convolutional neural network.

[0016] In an embodiment of the present application, the expression of the loss function is: ; wherein, in the formula s ( λ ) is the ground - truth input dataset, PSF ( λ ) is the mathematical expression of the PSF transmission of different spectra, D ( λ, PSF(λ) ) is the decoding network, is the chip response signal input dataset, τ is the weight factor, P 0( λ ) is the non - zero value of the initial network setting.

[0017] In an embodiment of the present application, the thickness of the metal layer is A, and 10 nm ≤ A ≤ 200 nm.

[0018] In an embodiment of the present application, the photoelectric sensor is a CCD or a CMOS.

[0019] By adopting the above - mentioned technical solutions, the resolution and imaging speed of the hyperspectral imager can be effectively improved. By combining spectral encoding and digital image reconstruction, the system can achieve fast, high - precision hyperspectral imaging of the target scene. Especially in the design of the spectral encoding module of the micro - nano structure array, the periodic distribution parameters match the characteristics of the point - spread function of the optical imaging component, which can greatly improve the extraction accuracy of spectral information, thereby improving the quality of the reconstructed spectral image. At the same time, the spectral reconstruction module is based on a deep - learning algorithm, ensuring robustness and efficiency under various environmental and noise conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] The present invention will be described in detail below with reference to specific embodiments and the accompanying drawings, wherein: Figure 1 Schematic structural diagram of the first embodiment of the present invention; Figure 2 is Figure 1 Amplified structural diagram of the spectral encoding module in; Figure 3 is Figure 2 Amplified structural diagram of the nano - photonic aperture array in; 10. Optical imaging component; 20. Spectral encoding module; 30. Photoelectric sensor; 21. Nano - photonic aperture array. Detailed implementation manners

[0021] In order to make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the following specific embodiments are only used to explain the present invention and do not constitute a limitation to the present invention.

[0022] As Figures 1 to 3 shown, in order to achieve the above - mentioned objectives, the present invention proposes a novel on - chip encoded - computing hyperspectral imager, including: An optical imaging component 10, configured to spatially focus the light rays of the target scene to form a two - dimensional light field; A spectral encoding module 20, which modulates the spectral information of the two - dimensional light field through a periodic micro - nano structure array to generate an encoded optical signal; A photoelectric sensor 30, configured to convert the encoded optical signal into digital image data; A spectral reconstruction module, which decodes the digital image data based on a deep learning algorithm to generate a three - dimensional data cube containing spatial - spectral information; Wherein, the periodic distribution parameters of the micro - nano structure array match the characteristics of the point spread function of the optical imaging component 10.

[0023] Specifically, the optical imaging component 10 includes a small optical lens and an optical system cooperating therewith. The small optical lens is responsible for capturing the light rays of the target scene and focusing them through optical elements such as lenses, mirrors, etc. The focused light rays form a two - dimensional light field and are transmitted to the subsequent spectral encoding module 20. The design of the optical imaging component 10 is optimized according to the resolution requirements of the target scene to ensure that the optical imaging accuracy and resolution meet the requirements.

[0024] The spectral encoding module 20 consists of a micro-nano structure array, which is used to modulate the spectral information of the two-dimensional light field generated by the optical imaging component 10 to generate an encoded optical signal. The specific structure of the micro-nano structure array is a periodically arranged nano-photonic hole array 21, and parameters such as the aperture, the arrangement period of the hole positions, and the shape are precisely designed to achieve effective encoding of different spectral information. The periodic distribution parameters of the micro-nano structure array match the characteristics of the point spread function of the optical imaging component 10 to ensure accurate modulation and encoding of spectral information during the encoding process. The material of this micro-nano structure array is a high-precision glass substrate, which is fabricated on a UV-curable polymer thin film through lithography and imprinting techniques, and a metal layer is deposited on the thin film to complete the final spectral encoding function.

[0025] The optoelectronic sensor 30 includes a CCD or CMOS sensor module, which is responsible for receiving the encoded optical signal processed by the spectral encoding module 20. The optoelectronic sensor 30 converts the optical signal into an electrical signal through optoelectronic conversion and generates digital image data. The digital image data contains the combination of spectral information and two-dimensional spatial imaging information generated by the spectral encoding module 20, and these data are transmitted to the spectral reconstruction module for subsequent processing.

[0026] The spectral reconstruction module is based on deep learning algorithms, mainly using a convolutional neural network (CNN) to decode the digital image data. The specific process includes using the trained network model to restore the input encoded image data to complete spectral information. During the spectral reconstruction process, the network will accurately reconstruct the spatial and spectral information through an adaptive deep learning algorithm, thereby generating a three-dimensional data cube containing spatial and spectral information. This module combines the point spread function (PSF) model in the optical imaging process to optimize the decoding effect and ensure the high-precision reconstruction of the three-dimensional data cube.

[0027] The optical imaging component 10 focuses the light of the target scene and forms a two-dimensional light field, and the spectral encoding module 20 modulates the spectral information of this light field through its micro-nano structure array to generate an encoded optical signal.

[0028] The encoded optical signal is received by the optoelectronic sensor 30 module. The optoelectronic sensor 30 receives the encoded optical signal and converts it into an electrical signal, generates digital image data, and sends it to the spectral reconstruction module for decoding.

[0029] The optoelectronic sensor 30 transmits the generated digital image data to the spectral reconstruction module. The spectral reconstruction module processes these data using deep learning algorithms to generate a high-precision three-dimensional data cube.

[0030] The spectral reconstruction module independently completes the decoding process of the digital image data and generates the final three-dimensional data cube output, which contains the spatial information and spectral information of the target scene.

[0031] By adopting the above technical solution, the resolution and imaging speed of the hyperspectral imager can be effectively improved. By combining spectral encoding and digital image reconstruction, the system can achieve fast, high-precision hyperspectral imaging of the target scene. Especially in the design of the spectral encoding module 20 of the micro-nano structure array, the periodic distribution parameters match the characteristics of the point spread function of the optical imaging component 10, which can greatly improve the extraction accuracy of spectral information, thereby improving the quality of the reconstructed spectral image. At the same time, the spectral reconstruction module is based on a deep learning algorithm, ensuring robustness and efficiency under various environmental and noise conditions.

[0032] In an embodiment of the present application, the spectral encoding module 20 includes: A transparent glass substrate; An ultraviolet curable polymer thin film disposed on the transparent glass substrate, and a structure array composed of a plurality of nano-photonic hole arrays 21 is formed on the ultraviolet curable polymer thin film; and A metal layer deposited on the side of the ultraviolet curable polymer thin film away from the transparent glass substrate.

[0033] Specifically, the transparent glass substrate is the support substrate of the entire spectral encoding module 20, and its function is to provide physical support for the subsequent structure and ensure the optical transmission performance of the system. The glass substrate is made of high-precision glass material to ensure that it does not cause obvious optical distortion during spectral encoding and guarantee the high imaging quality of the entire system. The thickness and size of the glass sheet are optimized according to the requirements of the optical system to ensure its mechanical strength and optical performance.

[0034] The ultraviolet curable polymer thin film is disposed on the transparent glass substrate. The thin film uses ultraviolet curable material and is cured into a stable structure by ultraviolet irradiation during the manufacturing process. The thickness and material properties of the polymer thin film determine the stability and accuracy of the micro-nano structure array. A plurality of nano-photonic hole arrays 21 are formed on the surface of the ultraviolet curable polymer thin film through photolithography and imprinting processes. This process can accurately control the size, shape and arrangement period of each aperture, thereby realizing efficient encoding of spectral information.

[0035] On the surface of the ultraviolet curable polymer thin film, a plurality of nano-photonic hole arrays 21 formed by photolithography or imprinting technology are the core part of the spectral encoding module 20. The design of the aperture size, shape and arrangement period of each nano-photonic hole will determine the accuracy and efficiency of spectral encoding. Specifically, the aperture of the nano-photonic hole can be distributed between 100 nm and 850 nm, and there are various choices for the hole position arrangement period and shape, such as square, regular hexagon, etc. Through these microstructures, the spectral encoding module 20 can perform spectral modulation on the passing optical signal.

[0036] The metal layer is deposited on the side of the ultraviolet-curable polymer thin film away from the transparent glass substrate, mainly enhancing optical properties and providing an optical reflection function. The thickness of the metal layer is controlled between 10 nm and 200 nm. Generally, metals with good optical reflection properties, such as aluminum, gold, silver, etc., are selected as the metal material. Vacuum evaporation, sputtering and other techniques are used in the deposition process of the metal layer to ensure that the metal layer uniformly covers the entire surface, so as to ensure the stability of spectral encoding.

[0037] The transparent glass substrate provides support for the spectral encoding module 20. The ultraviolet-curable polymer thin film is fixed on the glass substrate by bonding or other physical connection methods to ensure the stability and structural strength of the optical components. The nano-photonic pore array 21 structure formed on the surface of the polymer thin film by photolithography and imprinting techniques is directly located above the transparent glass substrate.

[0038] The metal layer is deposited on the upper surface of the ultraviolet-curable polymer thin film, on the side away from the transparent glass substrate. The role of the metal layer is to enhance the reflection and modulation efficiency of optical signals, and cooperate with the optical properties of the nano-photonic pore array 21 to ensure the accuracy and stability of spectral encoding.

[0039] Adopting the above technical solution, by forming the nano-photonic pore array 21 on the ultraviolet-curable polymer thin film and depositing a metal layer above it, high-precision spectral encoding can be achieved. The transparent glass substrate provides stable support, ensuring the mechanical and optical stability of the entire structure. This structural design can effectively utilize photolithography and imprinting techniques to precisely control the formation of micro-nano structures, with high manufacturing precision and reliability.

[0040] In an embodiment of the present application, the structure array is a 16*16 rectangle, and the pore size and / or the pore position arrangement period and / or the pore position arrangement shape in each nano-photonic pore array 21 are different.

[0041] Specifically, the overall structure array is in the form of a 16×16 rectangle, that is, there are a total of 16 rows and 16 columns of array units. Each array unit contains a nano-photonic pore array 21. The arrangement structure of the array is a rectangular array to ensure uniform distribution and sufficient spacing to optimize the effect of spectral encoding. The arrangement of the array is regular and symmetric on the entire spectral encoding module 20, which helps to efficiently achieve spectral encoding.

[0042] The pore size, the pore position arrangement period and the pore position arrangement shape in each nano-photonic pore array 21 are different. Specifically: The pore size of each nano-photonic pore can vary between 100 nm and 850 nm. This design enables the array to process spectral information of different wavelengths.

[0043] The hole position arrangement period of each array is between 300 nm and 900 nm, which is selected according to optical requirements during design to ensure the uniqueness of spectral modulation between different arrays and avoid spectral overlap or interference.

[0044] The hole position arrangement shapes within different arrays may be square, regular hexagon, etc. The variation in shapes helps to perform effective spectral modulation within different spectral ranges. This variation can improve the resolution and coding efficiency of spectral information.

[0045] Adopting the above technical solution, the 16×16 rectangular array structure of the spectral encoding module 20 enables more efficient and uniform modulation of spectral information. The diverse designs of the aperture size, hole position arrangement period, and hole position arrangement shape of each nano - photonic hole array 21 ensure high - precision modulation and uniqueness of different spectral information, which helps to achieve fine spectral encoding within multiple wavelength ranges.

[0046] In an embodiment of the present application, the aperture size within each nano - photonic hole array 21 falls within a first preset range, and the first preset range is from 100 nm to 850 nm.

[0047] Adopting the above technical solution, by controlling the aperture size within each nano - photonic hole array 21 within the first preset range of 100 nm to 850 nm, efficient and accurate spectral encoding can be ensured. This aperture range enables the spectral encoding module 20 to modulate within multiple wavelength ranges, ensuring that optical signals in different bands can be accurately encoded and processed.

[0048] In an embodiment of the present application, the hole position arrangement period within each nano - photonic hole array 21 falls within a second preset range, and the second preset range is from 300 nm to 900 nm.

[0049] Adopting the above technical solution, the hole position arrangement period within all nano - photonic hole arrays 21 falls within the second preset range of 300 nm to 900 nm, enabling the spectral encoding module 20 to effectively modulate optical signals in different spectral bands. The design of the hole position arrangement period within this range can ensure good modulation effect and high resolution of spectral encoding within multiple spectral band ranges.

[0050] In an embodiment of the present application, the arrangement shape of each nano - photonic hole array 21 is square or hexagonal.

[0051] With the above technical solution, by selecting the arrangement shape of a square or a hexagon, the square arrangement helps to achieve regular and symmetric spectral modulation, while the hexagon arrangement is more optimized in terms of space utilization efficiency, can densely arrange the hole positions, and provides higher spectral resolution and modulation density. The combination of the two arrangement shapes enables the spectral encoding module 20 to meet the high-precision modulation requirements within different wavelength ranges and improve the accuracy and resolution of the spectral image.

[0052] In an embodiment of the present application, the deep learning architecture adopts a CNN basic architecture. Using the response data set generated by the target radiation spectrum irradiated by a standard light source as the training input, and the response data set generated by a standard spectrometer receiving the same spectrum as the ground truth input, a loss function that embeds the knowledge of the optical PSF with the mean square error value of the spectral difference is used to train the spectral restoration convolutional neural network.

[0053] Specifically, the spectral image reconstruction system uses a convolutional neural network (CNN) as the basic architecture of deep learning. The CNN architecture extracts and processes the input data through structures such as convolutional layers, pooling layers, and fully connected layers, and is suitable for processing image data with spatial locality. During the spectral image reconstruction process, the CNN can effectively learn the correlation of spatial-spectral information from the training data and achieve efficient spectral reconstruction.

[0054] The radiation spectrum signal generated when the standard light source irradiates the target is modulated by the spectral encoding module 20 and a response data set is generated as the training input. This response data set contains the spectral information and imaging information of the target and is input into the convolutional neural network together with other image data for training. The radiation spectrum of the standard light source can cover multiple spectral bands, providing rich spectral features for the CNN to learn and reconstruct the spectral image of the target.

[0055] The standard spectrometer receives the same spectral signal as the target radiation spectrum and generates a corresponding response data set as the ground truth input data set. The output data set of the standard spectrometer is a real spectral image with high spectral accuracy and serves as the target output data during network training.

[0056] During the training process, the mean square error value of the spectral difference is used as the main loss function to measure the difference between the CNN reconstructed image and the ground truth image. In addition, the knowledge of the optical point spread function (PSF) is embedded in the loss function. The loss function not only considers the accuracy of image reconstruction but also includes the imaging characteristics of the optical system, that is, the influence of the PSF. By integrating the PSF knowledge into the loss function, the network can better simulate the blurring effect during the optical imaging process and improve the accuracy of image reconstruction.

[0057] Using the response data set generated by irradiating the target with a standard light source as the input, and the response data set generated by a standard spectrometer as the ground truth input, the convolutional neural network is trained by the backpropagation algorithm. During the training process, the network continuously adjusts the weights of the convolutional layers, reduces the value of the loss function, and optimizes the reconstruction effect of the spectral image.

[0058] After the training is completed, the convolutional neural network can output a high-precision spectral image similar to the measurement data of the standard spectrometer according to the input encoded optical signal data. These spectral images contain the three-dimensional spatial-spectral information of the target, that is, both two-dimensional spatial information and spectral information at each spatial position, forming the final hyperspectral data cube.

[0059] Adopting the above technical solution, the deep learning architecture adopts the basic architecture of a convolutional neural network (CNN), which can effectively learn from the response data set generated by the target radiation spectrum under the irradiation of a standard light source and the ground truth data of the standard spectrometer, and realize the efficient reconstruction of the spectral image. By embedding the knowledge of the optical point spread function (PSF) into the loss function, the deep learning network can more accurately simulate and compensate for the blurring effect generated during the optical imaging process, thereby improving the accuracy and quality of the spectral image.

[0060] In an embodiment of the present application, the expression of the loss function is: ; wherein, in the formula s ( λ ) is the ground truth input data set, PSF ( λ ) is the mathematical expression of the PSF transmission of different spectra, D ( λ, PSF(λ) ) is the decoding network, is the chip response signal input data set, τ is the weight factor, P 0( λ ) is the non-zero value of the initial network setting.

[0061] Adopting the above loss function expression, by embedding the knowledge of the point spread function (PSF), the blurring effect during the optical imaging process can be more realistically simulated during the training of the convolutional neural network, thereby effectively improving the accuracy and stability of the spectral image reconstruction. The mean square error term in the loss function is used to minimize the error between the network output and the ground truth data, while the embedding of the PSF further optimizes the network's adaptability to the characteristics of the optical system. By adjusting the weight factor τ, the balance between the reconstruction accuracy and regularization can be flexibly controlled to ensure that the network can converge faster and improve the reconstruction quality.

[0062] In an embodiment of the present application, the thickness of the metal layer is A, where 10 nm ≤ A ≤ 200 nm.

[0063] By adopting the above technical solution, the thickness of the metal layer is controlled between 10 nm and 200 nm, which can effectively improve the reflection efficiency of the spectral encoding module 20 and enhance its modulation ability for optical signals.

[0064] In an embodiment of the present application, the photoelectric sensor 30 is a CCD or a CMOS.

[0065] Specifically, the photoelectric sensor 30 can be a CCD (Charge-Coupled Device) or a CMOS (Complementary Metal-Oxide-Semiconductor) sensor. Both are commonly used image sensors that can convert the received optical signals into electrical signals and generate digital image data.

[0066] The CCD sensor has high imaging quality and low noise level, and is suitable for application scenarios with high requirements for image quality. The CCD sensor performs better in low-light environments and has good photoelectric conversion efficiency and uniformity.

[0067] Compared with the CCD, the CMOS sensor has lower power consumption and faster readout speed, and is suitable for applications with high requirements for speed and power consumption. The CMOS sensor can collect data at a higher frame rate and is suitable for high-speed imaging and real-time processing. Since both the CCD and the CMOS are prior arts, their working principles will not be elaborated in detail here.

[0068] By adopting the above technical solution, by reasonably selecting a CCD or a CMOS sensor, optimized image acquisition performance can be provided for the spectral image reconstruction system, ensuring the efficiency and adaptability of the system in different imaging tasks. This flexible selection of the photoelectric sensor 30 enables the spectral image reconstruction system to be widely applied to various hyperspectral imaging tasks.

[0069] The above are only the preferred embodiments of the present invention, and do not limit the patent scope of the present invention accordingly. Any equivalent structural transformation made under the inventive concept of the present invention by using the content of the specification and drawings of the present invention, or any direct / indirect application in other related technical fields, is included in the patent protection scope of the present invention.

Claims

1. A novel on-chip encoded computational hyperspectral imager, characterized in that, Comprising: An optical imaging component for spatially focusing the light of a target scene to form a two-dimensional light field; A spectral encoding module that modulates the spectral information of the two-dimensional light field through a periodic micro-nano structure array to generate an encoded optical signal; A photoelectric sensor for converting the encoded optical signal into digital image data; A spectral reconstruction module that decodes the digital image data based on a deep learning algorithm to generate a three-dimensional data cube containing spatial-spectral information; Wherein, the periodic distribution parameter of the micro-nano structure array matches the point spread function characteristics of the optical imaging component.

2. The novel on-chip encoded computational hyperspectral imager according to claim 1, characterized in that The spectral encoding module includes: A transparent glass substrate; An ultraviolet curable polymer film disposed on the transparent glass substrate, and a structure array composed of a plurality of nano-photonic hole arrays is formed on the ultraviolet curable polymer film; and A metal layer deposited on the side of the ultraviolet curable polymer film away from the transparent glass substrate.

3. The novel on-chip encoded computational hyperspectral imager according to claim 2, wherein The structure array is a 16*16 rectangle, and the pore size and / or the pore position arrangement period and / or the pore position arrangement shape in each nano-photonic hole array are different.

4. The novel on-chip encoded computational hyperspectral imager according to claim 3, wherein The pore size in each nano-photonic hole array falls within a first preset range, and the first preset range is from 100 nm to 850 nm.

5. The novel on-chip encoded computing hyperspectral imager according to claim 3, characterized in that, The pore position arrangement period in each nano-photonic hole array falls within a second preset range, and the second preset range is from 300 nm to 900 nm.

6. The novel on-chip encoded computational hyperspectral imager according to claim 3, characterized in that, The arrangement shape of each nano-photonic hole array is a square or a hexagon.

7. The novel on-chip coded computational hyperspectral imager according to claim 1, characterized in that, The deep learning architecture adopts the basic CNN architecture, uses the response dataset generated by the target radiation spectrum irradiated by a standard light source as the training input, uses the response dataset generated by a standard spectrometer receiving the same spectrum as the standard truth input, and embeds the mean square error value of the spectral difference between the two into the loss function of the optical PSF knowledge to train the spectral restoration convolutional neural network.

8. The novel on-chip encoded computational hyperspectral imager according to claim 5, wherein, The expression of the loss function is: ； Among them, in the formula s ( λ ) is the true value input data set, PSF ( λ ) is the mathematical expression of PSF transmission with different spectra, D ( λ,PSF (λ) ) is the decoding network, is the chip response signal input data set, τ is the weight factor, P 0( λ ) is the non-zero value of the initial network setting.

9. The novel on-chip encoded computational hyperspectral imager according to claim 2, wherein The thickness of the metal layer is A, and 10 nm ≤ A ≤ 200 nm.

10. The novel on-chip encoded computational hyperspectral imager according to claim 1, characterized in that, The photoelectric sensor is a CCD or a CMOS.

Images