Hyperspectral coding and lightweight classification and recognition method based on van der waals heterojunction

Abstract

The present application relates to a kind of high spectral coding based on van der waals heterojunction and light weight classification identification method, the method includes the inner product operation of received hyperspectral data with response matrix, the encoding feature set of device, response matrix is the response matrix of device calibration;Filter out multiple target encoding features in encoding feature set;Target encoding feature is input light weight classification model, and light weight classification model exports the pixel classification result corresponding to hyperspectral data.In the technical scheme of the present application, spectrum acquisition, coding is cooperated in device front end, realizes the cooperation between acquisition strategy of full waveband coverage and the sparse feature demand of downstream task, effectively reduces the unnecessary acquisition and transmission of redundant waveband.

Description

Technical Field

[0001] This invention relates to the field of optoelectronic technology, and in particular to a hyperspectral coding and lightweight classification and identification method based on van der Waals heterojunctions. Background Technology

[0002] Hyperspectral imaging technology can provide continuous spectral information for each pixel, offering richer spectral features compared to red, green, and blue three-primary-color imaging. However, existing hyperspectral sensing systems generally follow a sequential paradigm of "complete spectrum acquisition—digital processing—task inference," meaning that first, complete spectral information for each pixel in the scene is acquired using an imaging spectrometer to construct a high-dimensional data cube, followed by preprocessing, dimensionality reduction, and subsequent digital processing. However, the lack of coordination between the full-band coverage acquisition strategy and the sparse feature requirements of downstream tasks leads to the unnecessary acquisition and transmission of a large number of redundant bands. Summary of the Invention

[0003] In view of this, the present invention provides a hyperspectral coding and lightweight classification and identification method based on van der Waals heterojunctions, which achieves synergy between the acquisition strategy with full-band coverage and the sparse feature requirements of downstream tasks, effectively reducing unnecessary acquisition and transmission of redundant bands. The technical solution is as follows.

[0004] On the one hand, a hyperspectral coding and lightweight classification and recognition method based on van der Waals heterostructures is provided, the method comprising: The received hyperspectral data is multiplied by the response matrix to generate the device's coded feature set, and the response matrix is ​​the device's calibrated response matrix. Filter out multiple target coding features from the coding feature set; The target encoded features are input into a lightweight classification model, which outputs the pixel classification results corresponding to the hyperspectral data.

[0005] Furthermore, the device includes: an insulating layer, a photosensitive channel layer, a light-absorbing layer, and a bottom electrode; wherein, the bottom electrode is disposed on the surface of the insulating layer, a portion of the photosensitive channel layer covers the bottom electrode, another portion of the photosensitive channel layer covers and exposes the surface of the insulating layer, and the light-absorbing layer is at least partially disposed on the surface of the photosensitive channel layer; the light-absorbing layer and the photosensitive channel layer form a vertical van der Waals heterojunction.

[0006] Furthermore, the wavelength range of the response of the photosensitive channel layer is 450nm-1700nm.

[0007] Furthermore, the photosensitive channel layer is made of molybdenum distellide (MoTe2); the light-absorbing layer is made of palladium diselenide (PdSe2); and the insulating layer is made of silicon dioxide (SiO2).

[0008] Furthermore, the thickness of the photosensitive channel layer ranges from 8 nm to 15 nm.

[0009] Furthermore, the thickness of the light-absorbing layer ranges from 90nm to 130nm.

[0010] Furthermore, the bottom electrode includes a source electrode and a drain electrode, and the device is fabricated in the following manner: The source electrode and the drain electrode are formed on the insulating layer. The photosensitive channel layer MoTe2 is transferred between the source electrode and the drain electrode. The light-absorbing layer of PdSe2 is stacked on the photosensitive channel layer MoTe2 to form a vertical van der Waals heterojunction.

[0011] Furthermore, before generating the device's coded feature set, the following steps are also included: Input monochromatic light of different wavelengths into the device; Determine the photocurrent output by the device at each gate voltage; Based on the incident light intensity and the photocurrent at different wavelengths and gate voltages, the calibrated response matrix is ​​obtained. The response matrix is ​​determined by the spectral response function corresponding to each gate voltage.

[0012] Furthermore, multiple target coding features are selected from the coding feature set, including: Multiple target coding features are selected from the coding feature set by using orthogonal subspace projection or attention mechanism.

[0013] Furthermore, lightweight classification models include lightweight multilayer perceptrons (MLPs).

[0014] The hyperspectral coding and lightweight classification method based on van der Waals heterojunctions provided by this invention has at least the following beneficial effects: Since each device has a calibrated response matrix, meaning it corresponds to a unique response matrix, the inner product operation between the hyperspectral data and the response matrix yields the device's coded feature set. Multiple target coded features are then selected from this set and input into a lightweight classification model, outputting the pixel classification results corresponding to the hyperspectral data. Thus, spectral acquisition and encoding are collaboratively completed at the device's front end, achieving synergy between a full-band coverage acquisition strategy and the sparse feature requirements of downstream tasks, effectively reducing unnecessary acquisition and transmission of redundant bands.

[0015] In addition, the device utilizes the electrostatic shielding effect caused by the bottom electrode to form a reversible lateral homojunction in the photosensitive channel layer of the device, realizing gate voltage controllable switching of the bipolar photovoltaic response polarity under zero bias voltage; different gate voltages correspond to different spectral response functions, forming a programmable spectral response set. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other embodiments can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a schematic diagram of the structure of a sensor provided in one embodiment of the present invention; Figure 2 This is a schematic diagram of a sensor with photocurrents of opposite polarities under positive and negative grid voltages, provided in one embodiment of the present invention. Figure 3 This is a flowchart illustrating a hyperspectral coding and lightweight classification and identification method based on van der Waals heterojunctions, as provided in one embodiment of the present invention. Figure 4 A flowchart illustrating another hyperspectral coding and lightweight classification and identification method based on van der Waals heterojunctions is provided for one embodiment of the present invention. Figure 5 This is a schematic diagram illustrating the ability of a filtering channel to reduce data volume, as provided in one embodiment of the present invention. Figure 6 A schematic diagram illustrating the power reduction capability of a filtering channel according to an embodiment of the present invention; Figure 7 A schematic diagram illustrating a sensor with single-peak reconstruction capability according to an embodiment of the present invention; Figure 8 A schematic diagram illustrating a sensor with bimodal reconstruction capability according to an embodiment of the present invention; Figure 9 This is a flowchart illustrating another hyperspectral coding and lightweight classification and identification method based on van der Waals heterojunctions, provided as an embodiment of the present invention.

[0018] [Explanation of Labels in the Attached Image] 1: Substrate; 2: MoTe2 photosensitive channel layer; 3: Bottom electrode; 4: SiO2 insulating layer; 5: PdSe2 light-absorbing layer. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be further described in detail below with reference to specific examples and the accompanying drawings.

[0020] It should be noted that all uses of "first" and "second" in the embodiments of the present invention are for the purpose of distinguishing two entities or parameters with the same name but different names. It is clear that "first" and "second" are only for the convenience of expression and should not be construed as limiting the embodiments of the present invention. Subsequent embodiments will not explain this in detail.

[0021] This invention provides a sensor comprising a substrate, an insulating layer, a photosensitive channel layer, a light-absorbing layer, and a bottom electrode. The photosensitive channel layer is disposed on the substrate. The bottom electrode is disposed on the surface of the insulating layer. A portion of the photosensitive channel layer covers the bottom electrode, and another portion of the photosensitive channel layer covers the surface of the insulating layer. The light-absorbing layer is at least partially disposed on the surface of the photosensitive channel layer; the light-absorbing layer and the photosensitive channel layer form a vertical van der Waals heterojunction. It is evident that by utilizing the electrostatic shielding effect caused by the bottom electrode, the device eliminates static dark current power consumption in zero-bias operation mode, laying the foundation for low-power edge deployment from a hardware physical layer perspective. Furthermore, a reversible lateral homojunction is formed in the photosensitive channel layer, enabling controllable switching of the gate voltage for the bipolar photovoltaic response polarity under zero bias; different gate voltages correspond to different spectral response functions, constituting a programmable spectral response set.

[0022] The photosensitive channel layer has a response wavelength range of 450nm-1700nm. The thickness of the photosensitive channel layer ranges from 8nm to 15nm. The thickness of the light-absorbing layer ranges from 90nm to 130nm. The photosensitive channel layer is made of molybdenum distearate (MoTe2); the light-absorbing layer is made of palladium diselenide (PdSe2); and the insulating layer is made of silicon dioxide (SiO2). The bottom electrode includes a source electrode and a drain electrode. The device is fabricated as follows: the source electrode and drain electrode are formed on the insulating layer; the MoTe2 photosensitive channel layer is transferred between the source electrode and the drain electrode (referred to as the source-drain electrode); and the PdSe2 light-absorbing layer is stacked on top of the MoTe2 to form a vertical van der Waals heterojunction.

[0023] For example, Figure 1 This is a schematic diagram of the structure of a sensor provided in an embodiment of the present invention, as shown below. Figure 1As shown, this sensor can operate at 0V bias and uses its own power generation to power and regulate its own operation, i.e., self-driven photovoltaic regulation. The sensor may include a substrate 1, which can be made of silicon (Si). A MoTe2 photosensitive channel layer 2 is disposed on the substrate 1. A portion of the MoTe2 photosensitive channel layer 2 covers the bottom electrode 3, and another portion covers the surface of the SiO2 insulating layer 4. A PdSe2 light-absorbing layer 5 forms a vertical van der Waals heterojunction with the MoTe2 photosensitive channel layer 2. Of course, the MoTe2 not covered by PdSe2 can also be a light-absorbing layer; this is not specifically limited in this embodiment. Therefore, using a supercontinuum laser (SuperK Extreme, NKT Photonics) as the excitation source, the response was tested at room temperature with no source / drain bias, and with positive and negative gate voltages applied. Figure 2 As shown, the sensor outputs photocurrent under positive and negative grid voltages. For example, under a positive grid voltage (V=20V), the photocurrent is positive. Under a negative grid voltage (V=-20V), the photocurrent is negative.

[0024] The sensor fabrication process is described in detail below. The sensor fabrication process is as follows: Step 1: Bottom electrode fabrication. For example, the electrode pattern is defined on the SiO2 surface using standard photolithography: photoresist (RN246) is coated on a heavily doped Si substrate covered with a 285 nm SiO2 insulating layer; pre-baking at 90°C for 90 s; UV exposure (SUSS MJB4); post-baking at 90°C for 60 s; development with developer (NMD-32.38%) for 18 s; thermal evaporation (HHV FL400) to sequentially deposit Cr (10 nm) and Au (50 nm); acetone ultrasonic stripping to remove excess metal, rinsing with deionized water, and drying with nitrogen. The resulting Cr / Au electrode serves as the source / drain electrode and the non-uniform electrostatic shielding layer. Step 2: Two-dimensional material preparation and transfer. For example, MoTe2 exfoliation: a thin layer is obtained from a bulk MoTe2 crystal by mechanical exfoliation, and a sheet with a thickness of approximately 13.5 nm is picked up using a polydimethylsiloxane (PDMS) stamp; PdSe2 exfoliation: a PdSe2 sheet with a thickness of approximately 119.4 nm is obtained using the same method; dry transfer: with the assistance of an optical microscope, MoTe2 is first transferred between the source and drain electrodes, and then PdSe2 is stacked on top of MoTe2 to form a vertical heterojunction region. The transfer process is completed in a clean environment to avoid interface contamination. Step 3: Device characterization. Morphology: the device structure is observed using an optical microscope, and the thickness of each layer is measured using an atomic force microscope.

[0025] Based on the sensors provided in the above embodiments, this invention provides a hyperspectral coding and lightweight classification and identification method based on van der Waals heterojunctions. Figure 3A flowchart illustrating a hyperspectral coding and lightweight classification method based on van der Waals heterojunctions provided in this embodiment of the invention is shown below. Figure 3 As shown, the method may include: S301-S306 (some of which are optional).

[0026] S301, Inputting monochromatic light of different wavelengths into the device.

[0027] The device can be understood as the sensor in the above embodiments. Inputting monochromatic light of different wavelengths into the device can be understood as illuminating the light receiving end of the device with monochromatic light of different wavelengths. At this time, the device receives monochromatic light of different wavelengths.

[0028] S302. Determine the photocurrent output by the device at each gate voltage.

[0029] The device can adjust its gate voltage by being illuminated by monochromatic light of different wavelengths. Therefore, the device outputs photocurrent at different gate voltages.

[0030] S303. Based on the incident light intensity and the photocurrent at different wavelengths and different gate voltages, the calibrated response matrix is ​​obtained. The response matrix is ​​determined by the spectral response function corresponding to each gate voltage.

[0031] For example, S303 can be implemented such that the ratio of photocurrent to power is equal to the responsivity, where the power is determined based on the photocurrent and gate voltage. Thus, the responsivity forms the response matrix R(λ,Vg). Under illumination by unknown light, the photocurrent is recorded, and the spectral response function of the device at each gate voltage is calculated inversely based on this photocurrent and the response matrix.

[0032] The expressions for photocurrent, gate voltages, and response matrix can be:

[0033] in, Represented as photocurrent; It is represented as a response matrix, where rows represent different gate voltages and columns represent different wavelengths; It is expressed as the incident light intensity of monochromatic light of different wavelengths.

[0034] After obtaining the spectral response function of the device at each gate voltage, encoding is performed based on the device's unique hyperspectral response function, as described in step S304 below.

[0035] S304. Perform an inner product operation between the received hyperspectral data and the response matrix to generate the device's encoded feature set. The response matrix is ​​the device's calibrated response matrix.

[0036] The device's encoding feature set can be understood as a set of photocurrents, with each photocurrent represented as I(Vg).

[0037] For example, such as Figure 4 As shown, the hyperspectral data is illumination data with width (W) * height (H) * depth (N). After performing an inner product operation between the hyperspectral data and the response matrix, the device's encoded feature set is obtained, i.e., Figure 4 In this context, M represents M gate voltages. This means that the first image represents the image formed by the encoded features output by the device under one gate voltage. Similarly, the second image represents the image formed by the encoded features output by the device under another gate voltage. And so on, resulting in M ​​images.

[0038] S305. Select multiple target coding features from the coding feature set.

[0039] Using the example above, after obtaining M images, the target image is selected. Specifically, target coding features are selected from the coding feature set, for example, the second group of coding features, the fourth group of coding features, ... the Mth group of coding features are selected, and then these target coding features are input into the model in S306 for classification and recognition.

[0040] For example, orthogonal subspace projection (OSP) or attention mechanisms can be used to select multiple target encoded features from the encoded feature set. For instance, such as... Figure 5 and Figure 6 As shown, only 24 channels are needed to achieve classification performance close to the original 49-band complete spectrum, reducing the floating-point operation (FLOPs) of the back-end neural network by about 50%.

[0041] S306. Input the target encoding features into the lightweight classification model, and the lightweight classification model outputs the pixel classification results corresponding to the hyperspectral data.

[0042] Lightweight classification models can include lightweight multilayer perceptrons (MLPs).

[0043] Continuing with the above example, such as Figure 4As shown, the target encoded features are input into the MLP, and the output is a pixel classification result. The pixel classification result can include seven categories: cropland, forest, grassland, shrub, impervious surface, water, and bare land. Alternatively, the pixel classification result can also include seven categories: roofs, streets, paths, grass, trees, water, and shadows. Therefore, no specific limitation is made in this embodiment of the invention.

[0044] In this embodiment of the invention, the selected compact features are input to a lightweight multilayer perceptron (MLP) at the back end, and pixel-by-pixel classification and recognition are completed in zero-bias mode.

[0045] Of course, the sensor provided in the above embodiments uses a 10nm step size for response matrix calibration in the 640-1100nm wavelength range. Furthermore, by using algorithms to reconstruct unknown monochromatic light and bimodal spectra, it can be determined that the sensor has single-peak reconstruction capabilities (e.g., Figure 7 (as shown) and the sensor has bimodal reconstruction capability (such as...) Figure 8 (As shown). Therefore, in conjunction with the relevant descriptions of S304-S306 above, as follows: Figure 9 As shown, taking mixed light as an example, the encoding and pixel recognition of the mixed light beam can be found in the relevant descriptions above, and will not be repeated here.

[0046] In summary, since the device has a calibrated response matrix, it means that each device corresponds to a unique response matrix. Therefore, performing an inner product operation between the hyperspectral data and the response matrix yields the device's coded feature set. Multiple target coded features are then selected from this set and input into a lightweight classification model, outputting the pixel classification results corresponding to the hyperspectral data. It is evident that spectral acquisition and encoding are collaboratively completed at the device's front end, achieving synergy between a full-band coverage acquisition strategy and the sparse feature requirements of downstream tasks, effectively reducing unnecessary acquisition and transmission of redundant bands.

[0047] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0048] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A hyperspectral coding and lightweight classification and recognition method based on van der Waals heterojunctions, characterized in that, The method includes: The received hyperspectral data is multiplied by the response matrix to generate a set of encoded features for the device, wherein the response matrix is ​​the response matrix calibrated by the device. Select multiple target coding features from the set of coding features; The target encoded features are input into a lightweight classification model, which outputs the pixel classification results corresponding to the hyperspectral data.

2. The method according to claim 1, characterized in that, The device includes: an insulating layer, a photosensitive channel layer, a light-absorbing layer, and a bottom electrode; The bottom electrode is disposed on the surface of the insulating layer, a portion of the photosensitive channel layer covers the bottom electrode, another portion of the photosensitive channel layer covers the surface of the insulating layer, and the light-absorbing layer is at least partially disposed on the surface of the photosensitive channel layer; the light-absorbing layer and the photosensitive channel layer form a vertical van der Waals heterojunction.

3. The method according to claim 2, characterized in that, The wavelength range of the response of the photosensitive channel layer is 450nm-1700nm.

4. The method according to claim 2, characterized in that, The photosensitive channel layer is made of molybdenum distellide; the light-absorbing layer is made of palladium diselenide; and the insulating layer is made of silicon dioxide.

5. The method according to claim 2, characterized in that, The thickness of the photosensitive channel layer ranges from 8 nm to 15 nm.

6. The method according to claim 2, characterized in that, The thickness of the light-absorbing layer ranges from 90nm to 130nm.

7. The method according to claim 2, characterized in that, The bottom electrode includes a source electrode and a drain electrode, and the device is fabricated in the following manner: The source electrode and the drain electrode are formed on the insulating layer, the photosensitive channel layer is transferred between the source electrode and the drain electrode, and the light-absorbing layer is stacked on the photosensitive channel layer to form the vertical van der Waals heterojunction.

8. The method according to any one of claims 1-7, characterized in that, Before the set of encoded features of the generating device, it also includes: Monochromatic light of different wavelengths is input into the device; Determine the photocurrent output by the device at each gate voltage; The calibrated response matrix is ​​obtained based on the incident light intensity and the photocurrent at different wavelengths and different gate voltages. The response matrix is ​​determined by the spectral response function corresponding to each gate voltage.

9. The method according to any one of claims 1-7, characterized in that, The process of selecting multiple target coding features from the coding feature set includes: Multiple target coding features in the coding feature set are selected by using an orthogonal subspace projection method or an attention mechanism.

10. The method according to any one of claims 1-7, characterized in that, The lightweight classification model includes a lightweight multilayer perceptron (MLP).

Images