Intelligent design method of wide-band infrared polarization imaging device

Abstract

The application discloses an intelligent design method of a broadband infrared polarization imaging device, belongs to the technical field of deep learning and metamaterials, and can solve the problem that a traditional design method seriously depends on prior knowledge of researchers and uses an iterative trial-and-error method of full-wave simulation, thereby leading to a long and inefficient design process. An end-to-end bidirectional design framework is constructed, including two core modules of a forward design network and a reverse design network. The forward design network adopts a multi-task learning strategy, takes structural parameters of the device as input, and respectively outputs polarization responses of TM and TE light, so that multi-channel modeling and collaborative prediction are realized. The reverse design network utilizes a convolution network architecture, takes the polarization responses of TM and TE light as input, and outputs corresponding structural parameters, so that a high-dimensional nonlinear problem of "spectrum-to-structure" is effectively solved. The method can significantly improve the design precision and efficiency of the broadband infrared polarization imaging device.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of deep learning and metamaterials, and particularly relates to an intelligent design method of a wide-band infrared polarization imaging device. BACKGROUND

[0002] Infrared polarization imaging technology converts temperature and polarization information into images for display by taking advantage of the differences in infrared radiation and polarization characteristics of different targets. Since infrared imaging has higher sensitivity at night or in bad weather, and polarization can highlight the details and contours of the target, combining infrared and polarization technology can effectively distinguish artificial targets from natural backgrounds in complex environments, improve target recognition accuracy and detection efficiency, and achieve efficient and accurate target recognition. Therefore, infrared polarization imaging technology has been widely used in many fields such as vegetation coverage, remote sensing detection, target recognition and tracking.

[0003] With the increasing demand for infrared polarization imaging of dynamic targets, focal plane infrared polarization imaging technology has become a research hotspot in various countries. Subwavelength gratings (SWG) have significant advantages in miniaturization, integration, and polarization performance, and have the characteristics of real-time and high resolution. Therefore, many infrared subwavelength gratings have been proposed. Among them, the growth of a 13:1 high aspect ratio metal grating on the sidewall of a dielectric grating can achieve a transmittance of 70% and an extinction ratio of 29.1 dB within a wavelength of 1.05-1.6um. The SWG composed of metal-dielectric-metal (MIM) gratings can efficiently suppress polarized light in a specific direction due to the formation of a Fabry-Perot cavity by the MIM structure, thus having high polarization performance. The transmittance and extinction ratio reach 85% and 70dB respectively within a wavelength of 1.5-8um. The polarizer composed of an interlayer-metal (IM) double-layer grating structure can achieve a transmittance of 90.8% and an extinction ratio of 50dB within a wavelength of 7-15um.

[0004] However, traditional SWG design and optimization mainly rely on numerical simulation methods such as the Finite Difference Time Domain Method (FDTD) and the Finite Element Method (FEM). When optimizing multi-parameter designs over a wide spectral range, the computational time and resource requirements increase significantly, and the efficiency is low when facing large-scale design space searches. It heavily relies on researchers' prior knowledge and iterative trial-and-error methods using full-wave simulation, resulting in a lengthy and inefficient design process. Furthermore, it suffers from high-dimensional nonlinearity issues related to "inferring structure from spectrum." In addition, due to the significant difference between TM and TE transmittance values, existing methods struggle to simultaneously predict TM and TE transmittances that differ by several orders of magnitude. This weakens the optimization effect of the TE task, ultimately reducing the prediction accuracy of TE and affecting model training performance. Therefore, how to perform SWG design and optimization efficiently and accurately over a wide infrared range has become a hot research topic for those skilled in the art.

[0005] In summary, existing technologies suffer from a technical problem of relying heavily on researchers' prior knowledge and using iterative trial-and-error methods based on full-wave simulations, resulting in lengthy and inefficient design processes. Summary of the Invention

[0006] This invention solves the technical problem that existing technologies rely heavily on researchers' prior knowledge and iterative trial-and-error methods using full-wave simulation, resulting in lengthy and inefficient design processes.

[0007] The intelligent design method for a wideband infrared polarization imaging device according to the present invention includes the following steps:

[0008] Step 1: Obtain multiple structural parameters of the subwavelength metal grating;

[0009] Step 2: Preset a forward design network, train the preset forward design network using multiple structural parameters of the subwavelength metal grating to obtain the forward design network, and predict the multiple structural parameters of the subwavelength metal grating based on the forward design network to obtain the predicted spectrum.

[0010] Step 3: Preset the inverse design network, train the preset inverse design network using the predicted spectrum to obtain the inverse design network, and back-infer the predicted parameters based on the inverse design network.

[0011] Step 4: Design the subwavelength metal grating based on the predicted parameters to obtain the designed subwavelength metal grating.

[0012] Furthermore, in this embodiment of the invention, the multiple structural parameters of the subwavelength metal grating in step 1 are the interval d, metal width W, metal height H1, intermediate dielectric layer height H2, period P, dielectric height H at the air trench, and dielectric width s at the air trench.

[0013] Further, in the embodiment of the present application, the step 2 is to predict the plurality of structure parameters of the subwavelength metal grating based on the forward design network, specifically:

[0014] The plurality of structure parameters of the subwavelength metal grating are processed, and the processed plurality of structure parameters of the subwavelength metal grating are predicted by using the shared hidden layer and the subnetwork of the forward design network to obtain a predicted spectrum.

[0015] Further, in the embodiment of the present application, the step 2 is to predict the plurality of structure parameters of the subwavelength metal grating based on the forward design network, specifically:

[0016] The plurality of structure parameters of the subwavelength metal grating are processed, and the processed plurality of structure parameters of the subwavelength metal grating are predicted by using the shared hidden layer and the subnetwork of the forward design network to obtain a predicted spectrum.

[0017] Further, in the embodiment of the present application, the step 2 is to predict the plurality of structure parameters of the subwavelength metal grating based on the forward design network, specifically:

[0018] The plurality of structure parameters of the subwavelength metal grating are processed, and the processed plurality of structure parameters of the subwavelength metal grating are predicted by using the shared hidden layer and the subnetwork of the forward design network to obtain a predicted spectrum.

[0019] Further, in the embodiment of the present application, the forward design network adopts a loss function composed of a normalized MSE and a normalized Log-MSE.

[0020] Further, in the embodiment of the present application, the step 3 is to back-propagate the predicted spectrum based on the reverse design network, specifically:

[0021] The predicted spectrum is converted into 992-dimensional data by using the convolution module of the reverse design network, and the 992-dimensional data is reduced in dimension by using the full connection module of the reverse design network to output a predicted parameter.

[0022] Further, in the embodiment of the present application, the convolution module of the reverse design network is sequentially connected with a first convolution layer, a first pooling layer, a second convolution layer and a second pooling layer.

[0023] Further, in the embodiment of the present application, the reverse design network adopts an MSE loss function.

[0024] The sub-wavelength metal grating is realized based on any of the above-mentioned methods, the sub-wavelength metal grating adopts a double-metal-medium-metal structure, the double-metal-medium-metal is separated by using an air groove, the metal is aluminum, the medium is silicon oxide, the substrate is aluminum oxide, and the wavelength range is 1-12 mu m.

[0025] The present application solves the technical problem in the prior art that the iterative trial-and-error method seriously depends on the prior knowledge of researchers and uses full-wave simulation, resulting in a long and inefficient design process.

[0026] 1. The present application proposes an intelligent design method for a wide-band infrared polarization imaging device, and constructs an end-to-end bidirectional design framework, including two core modules of a forward design network and a reverse design network. The forward design network and the reverse design network are each completed by the same neural network framework during the training process, realizing end-to-end modeling without relying on manual intervention or external simulation software, suitable for realizing the prediction of structure parameters and the prediction of polarization response of a wide-band infrared sub-wavelength metal grating without traditional simulation calculation, and completing efficient and high-precision modeling and optimization design between the grating structure and performance.

[0027] 2. The present application proposes an intelligent design method for a wide-band infrared polarization imaging device, the forward design network adopts a multi-task learning strategy, takes the structure parameters of the device as input, and respectively outputs the polarization responses of TM light and TE light, realizing multi-channel modeling and collaborative prediction; the reverse design network uses a convolutional network architecture, takes the polarization responses of TM light and TE light as input, and outputs the corresponding structure parameters, effectively solving the high-dimensional nonlinear problem of "spectrum-based structure reverse", and reducing the calculation time and resource demand when optimizing multi-parameter design in a wide-band range, which can realize high-precision and high-efficiency prediction and reverse design of sub-wavelength metal grating structure parameters when facing large-scale design space search.

[0028] 3. The present application proposes an intelligent design method for a wide-band infrared polarization imaging device, the forward design network can simultaneously predict the TM and TE transmittances with orders of magnitude difference by using a loss function composed of normalized MSE and normalized Log-MSE, solving the technical problem that due to the large difference in orders of magnitude between TM transmittance and TE transmittance, if the same loss function is used, the gradient of TM task will dominate, weakening the optimization effect of TE task, finally reducing the prediction accuracy of TE, and thus affecting the model training performance. BRIEF DESCRIPTION OF DRAWINGS

[0029] The above and / or additional aspects and advantages of the present application will become apparent and more readily appreciated from the following description, taken in conjunction with the following drawings of which:

[0030] Figure 1 is an infrared subwavelength metal grating structure diagram according to the second embodiment;

[0031] Figure 2 is a forward design network design structure diagram in the end-to-end multi-task convolutional bidirectional deep learning design method according to the second embodiment;

[0032] Figure 3 is a reverse design network design structure diagram in the end-to-end multi-task convolutional bidirectional deep learning design method according to the fourth embodiment;

[0033] Figure 4 is a whole design structure diagram of the end-to-end multi-task convolutional bidirectional deep learning design method according to the fifth embodiment;

[0034] Figure 5 is a training and verification loss diagram of the forward design network in the end-to-end multi-task convolutional bidirectional deep learning design method according to the fifth embodiment;

[0035] Figure 6 is a prediction result diagram of the forward design network in the end-to-end multi-task convolutional bidirectional deep learning design method according to the fifth embodiment;

[0036] Figure 7 is a forward design network and FDTD method running time comparison diagram in the end-to-end multi-task convolutional bidirectional deep learning design method according to the fifth embodiment;

[0037] Figure 8 is a training and verification loss diagram of the reverse design network in the end-to-end multi-task convolutional bidirectional deep learning design method according to the fifth embodiment;

[0038] Figure 9 is a prediction result diagram of the reverse design network in the end-to-end multi-task convolutional bidirectional deep learning design method according to the fifth embodiment;

[0039] Figure 10 is a running time result diagram of the reverse design network in the end-to-end multi-task convolutional bidirectional deep learning design method according to the fifth embodiment. DETAILED DESCRIPTION

[0040] Various embodiments of the present application will be described hereinafter with reference to the accompanying drawings. The embodiments described by reference to the drawings are illustrative and are intended to explain the present application, and are not to be construed as limiting the present application.

[0041] Embodiment one. An intelligent design method of a wide-band infrared polarization imaging device, comprising the following steps:

[0042] Step 1, obtaining a plurality of structure parameters of the sub-wavelength metal grating;

[0043] Step 2, presetting a forward design network, training the preset forward design network with the plurality of structure parameters of the sub-wavelength metal grating to obtain the forward design network, predicting the 7-dimensional structure parameters of the sub-wavelength metal grating based on the forward design network to obtain a predicted spectrum;

[0044] Step 3, presetting a reverse design network, training the preset reverse design network with the predicted spectrum to obtain the reverse design network, and inversely predicting the predicted spectrum based on the reverse design network to obtain a predicted parameter;

[0045] Step 4, designing the sub-wavelength metal grating based on the predicted parameter to obtain a designed sub-wavelength metal grating.

[0046] The prior art has the technical problem of being seriously dependent on the prior knowledge of researchers and using an iterative trial-and-error method of full-wave simulation, resulting in a long and inefficient design process.

[0047] To solve the above technical problem, the embodiment provides an intelligent design method of a wide-band infrared polarization imaging device, wherein in step 2, the preset forward design network is trained with the plurality of structure parameters of the sub-wavelength metal grating, wherein the number of training rounds adopted in the training process of the forward design network is 800, and the learning rate is dynamically adjusted according to the number of current training rounds (Epoch), and the expression is .

[0048] In step 3, the preset reverse design network is trained with the predicted spectrum, and the number of training rounds adopted in the training process of the reverse design network is 600, and the learning rate is dynamically adjusted according to the number of current training rounds (Epoch), and the expression is .

[0049] Therefore, based on the trained forward design network and reverse design network, the embodiment realizes end-to-end modeling without relying on manual intervention or external simulation software, is suitable for realizing the prediction of structure parameters and polarization response of the wide-band infrared sub-wavelength metal grating without traditional simulation calculation, completing the efficient and high-precision modeling and optimization design between the grating structure and performance, and solving the technical problem of the prior art that is seriously dependent on the prior knowledge of researchers and uses an iterative trial-and-error method of full-wave simulation, resulting in a long and inefficient design process.

[0050] Embodiment two. The embodiment is a further limitation of the intelligent design method of a wide-band infrared polarization imaging device described in embodiment one. The plurality of structural parameters of the sub-wavelength metal grating in step 1 are respectively the interval d, the metal width W, the metal height H1, the intermediate dielectric layer height H2, the period P, the dielectric height H at the air groove, and the air groove width s.

[0051] The plurality of structural parameters of the sub-wavelength metal grating in step 2 are predicted based on the forward design network, and specifically:

[0052] The plurality of structural parameters of the sub-wavelength metal grating are processed, and the processed plurality of structural parameters of the sub-wavelength metal grating are converted into 200-dimensional data using a shared hidden layer. The sub-network includes two networks with the same structure. The 200-dimensional data is equally divided into two parts. The two networks of the sub-network are used to upgrade the data of one part respectively. TM light and TE light are respectively incident into the sub-wavelength metal grating. Based on the upgraded data, the corresponding TM transmittance and TE transmittance are outputted, and the predicted spectrum is obtained.

[0053] The sub-wavelength metal grating described in the embodiment is as shown in Figure 1 The sub-wavelength metal grating adopts a double-metal-dielectric-metal structure and is separated by an intermediate air groove. It has 7 structural parameter variables, and TM light and TE light are respectively incident to obtain the transmittance of the corresponding structural parameters.

[0054] In the forward design network, the input is the 7-dimensional structural parameters of the sub-wavelength metal grating, and the output is the 501×2-dimensional spectrum data of the TM (Transverse Magnetic) transmittance and TE (Transverse Electric) transmittance.

[0055] As shown in Figure 2 The end-to-end multi-task convolutional bidirectional deep learning design method based on the wide-band infrared sub-wavelength metal grating proposed in the embodiment adopts a multi-task learning strategy for the forward design network. The 7-dimensional structural parameters are finally outputted as the 501×2-dimensional TM transmittance and TE transmittance in the 1-12 μm wide wavelength range. The input sequentially passes through the shared hidden layer and the sub-network. The shared hidden layer neurons are 60 and 200 respectively. The 7-dimensional input is sequentially upgraded to 60-dimensional and 200-dimensional using the shared hidden layer. Then, the 200-dimensional data is equally divided and passed through the sub-network. The sub-network performs multi-task learning. The sub-network has the same network structure, and the neurons are 400, 600, and 800 respectively. The 100-dimensional data equally divided is sequentially upgraded to 400-dimensional, 600-dimensional, and 800-dimensional using the sub-network. Finally, the 501-dimensional TM transmittance and TE transmittance in the 1-12 μm wide wavelength range are respectively outputted.

[0056] As a further design of the present embodiment, the plurality of structure parameters of the sub-wavelength metal grating are normalized to ensure the stability and efficiency of the training, and the normalization method is as follows:

[0057] ;

[0058] In the formula, is a parameter vector of a certain dimension, is the mean of the vector, is the variance.

[0059] Embodiment three. The present embodiment is a further limitation of the intelligent design method of the wide-band infrared polarization imaging device of embodiment one or two, and the forward design network adopts a loss function combined by normalized MSE and normalized Log-MSE.

[0060] In the training process of the forward design network, due to the large difference in the order of magnitude of TM transmittance and TE transmittance, if the same loss function is used, the gradient of TM task will dominate, weaken the optimization effect of TE task, and affect the model training performance. Therefore, a new weighting mechanism is adopted in the present embodiment, and a loss function combined by normalized MSE and normalized Log-MSE is constructed to realize scale balance between different tasks, and the specific expression is as follows

[0061] ;

[0062] In the formula, and are the loss weights of TM and TE tasks respectively, and The expressions of and are as follows:

[0063] ;

[0064] ;

[0065] In the formula, and are the real TM transmittance and TE transmittance respectively, and are the predicted TM transmittance and TE transmittance respectively, is the total number of data.

[0066] As a further design of the present embodiment, is 0.9, is 0.1, is 7.

[0067] Embodiment four. The embodiment is a further limitation of the intelligent design method of the wide-band infrared polarization imaging device of embodiment one, wherein the step 3 of inversely predicting the predicted spectrum based on the inverse design network is specifically:

[0068] The 992-dimensional data is converted from the predicted spectrum by the convolution module of the inverse design network, and the 992-dimensional data is reduced in dimension by the full connection module of the inverse design network to output the predicted parameters.

[0069] The end-to-end multi-task convolutional bidirectional deep learning design method based on the wide-band infrared sub-wavelength metal grating is proposed in the embodiment, wherein the inverse design network adopts a convolution module to finally output 7-dimensional structure parameters from the input 501*2-dimensional TM transmittance and TE transmittance in the wavelength range of 1-12, as shown in the following formula: Figure 3 The input data sequentially passes through the convolution module and the full connection module. Firstly, the 501*2-dimensional input data is converted into 992*1-dimensional data by the convolution module. The convolution module adopts two groups of one-dimensional convolution layers and pooling layers. The first convolution layer inputs 501*2 and outputs 251*8. The first pooling layer inputs 251*8 and outputs 126*8. The second convolution layer inputs 126*8 and outputs 124*16. The second pooling layer inputs 124*16 and outputs 62*16. Then, the data is reduced in dimension by the full connection module. The number of neurons of the full connection module is 992, 500 and 200 in sequence. The full connection module flattens the 62*16-dimensional data of the convolution module into 992*1, and reduces the dimension to 500 and 200 in sequence, and finally outputs the 7-dimensional normalized structure parameters.

[0070] Embodiment five. The embodiment is a further limitation of the intelligent design method of the wide-band infrared polarization imaging device of embodiment one or four, wherein the inverse design network adopts an MSE loss function.

[0071] The MSE is adopted as the loss function for optimization, and the specific expression is as follows:

[0072] ;

[0073] In the formula, Y and Y' respectively represent the real structure parameters and the predicted structure parameters.

[0074] In summary, the intelligent design method of the wide-band infrared polarization imaging device is proposed in the embodiment, and the end-to-end multi-task convolutional bidirectional deep learning model is formed by combining the forward design network and the inverse design network, as shown in the following formula: Figure 4 ​​As shown, the forward design network receives multiple structure parameters as input, extracts parameter features through the convolutional neural network, and finally outputs the corresponding transmittance spectrum data in TM and TE polarization states for describing the transmission performance of the grating at different wavebands. The reverse design network module takes the target transmittance spectrum as input, extracts features and outputs a corresponding set of structure parameters after the regression network, which is used to realize the reverse prediction of the structure. The reverse design network module reverses the possible geometric parameter combination according to the required spectral response, which assists in realizing the rapid reverse design of the grating structure. The model can realize the prediction of the bidirectional mapping relationship between the structure parameters of the wide-band infrared sub-wavelength metal grating and its transmittance in TM and TE polarization states.

[0075] As shown in Figure 5 , it is the training and validation loss chart of the forward design network. In the first 100 rounds, the training loss and the validation loss both show a clear downward trend. As the training proceeds, the learning rate gradually decreases, and the loss value gradually tends to be stable, indicating that the model training has reached a good optimization effect and no overfitting phenomenon occurs. As shown in Figure 6 , it is the comparison result of the true value and the predicted value of the forward design network. The result shows that the difference between the predicted value and the true value of the forward design network model is small, which verifies the superiority of the model in accuracy and stability, indicating that the network can make high-precision prediction and has high reliability in practical application. As shown in Figure 7 , it is the running time comparison of the forward design network and the finite-difference time-domain (FDTD) method. The results show that the running time of the FDTD method is in the range of several seconds, specifically 1.82~7.26s. In contrast, the running time of the deep learning method is within milliseconds, and the time of most samples is within 1ms, which is about 0.0002 times of the FDTD method. This proves that the deep learning method has the advantage of high efficiency in the design of sub-wavelength metal gratings.

[0076] As shown in Figure 8 , it is the training and validation loss chart of the reverse design network. As can be seen from the chart, in the 600 rounds, the training loss and the validation loss both show a steady downward trend, indicating that the reverse design network has shown good optimization effect in the training process, the loss gradually decreases and no overfitting phenomenon occurs, ensuring that the network can effectively learn the accurate mapping relationship from the transmittance to the structure parameters. As shown in Figure 9 , it is the comparison result of the true value and the predicted value of the reverse design network. The result shows that the difference between the predicted value and the true value of the model is small, indicating the accuracy of the network in the reverse derivation of the structure parameters. These results verify the effectiveness of the network in the high-dimensional structure parameter space, which can provide reliable prediction in complex design tasks and has good robustness and precision. As shown in Figure 10The running time curve of the inverse design network is shown, and it can be seen from the figure that the running time of the inverse design network is 0.5ms-1.6ms, and mainly concentrates in 1ms, which is about 0.0002 times of the FDTD method. It is proved that the inverse design network can design the broadband infrared subwavelength metal grating with high efficiency and high precision.

[0077] Embodiment six. The subwavelength metal grating described in this embodiment is realized based on the method described in any one of embodiments one to five, the subwavelength metal grating adopts a double-metal-dielectric-metal structure, the double-metal-dielectric-metal is separated by air grooves, the metal is aluminum, the dielectric is silicon oxide, the substrate is aluminum oxide, and the wavelength range is 1-12 μm.

[0078] The intelligent design method of the broadband infrared polarization imaging device is described in detail above, and the principles and implementation modes of the present application are described by applying specific examples; the above examples are only used to help understand the method and core idea of the present application; meanwhile, for those skilled in the art, according to the idea of the present application, the specific implementation mode and application range will be changed, and the above description should not be understood as a limitation of the present application.

Claims

1. An intelligent design method for a broadband infrared polarization imaging device, characterized in that, Includes the following steps: Step 1: Obtain multiple structural parameters of the subwavelength metal grating; Step 2: Preset a forward design network, train the preset forward design network using multiple structural parameters of the subwavelength metal grating to obtain the forward design network, and predict the multiple structural parameters of the subwavelength metal grating based on the forward design network to obtain the predicted spectrum. Step 2, which involves predicting multiple structural parameters of the subwavelength metal grating using a forward design network, specifically includes: Multiple structural parameters of the subwavelength metal grating are processed, and the shared hidden layer and subnetwork of the forward-designed network are used to predict the multiple structural parameters of the processed subwavelength metal grating to obtain the predicted spectrum. The method of predicting multiple structural parameters of the processed subwavelength metal grating using the shared hidden layer and subnetwork of the forward-designed network is as follows: The shared hidden layer is used to convert multiple structural parameters of the processed subwavelength metal grating into 200-dimensional data, and the 200-dimensional data is input into the sub-network for prediction. The process of inputting 200-dimensional data into the sub-network for prediction specifically involves: The subnetwork consists of two networks with the same structure. The 200-dimensional data is divided into two parts. The two networks of the subnetwork are used to upgrade one part of the data. The TM light and TE light are respectively injected into the subwavelength metal grating. Based on the upgraded data, the corresponding TM transmittance and TE transmittance are output to obtain the predicted spectrum. Step 3: Preset the inverse design network, train the preset inverse design network using the predicted spectrum to obtain the inverse design network, and back-infer the predicted parameters based on the inverse design network. Step 3, which involves reverse-engineering the predicted spectrum using a reverse-designed network, specifically involves: The predicted spectrum is converted into 992-dimensional data using the convolutional module of the inverse design network. The dimensionality of the 992-dimensional data is reduced using the fully connected module of the inverse design network, and the predicted parameters are output. Step 4: Design the subwavelength metal grating based on the predicted parameters to obtain the designed subwavelength metal grating.

2. The intelligent design method for a wideband infrared polarization imaging device according to claim 1, characterized in that, The subwavelength metal grating in step 1 has several structural parameters, namely, the spacing d, the metal width W, the metal height H1, the height H2 of the intermediate dielectric layer, the period P, the dielectric height H at the air trench, and the dielectric width s at the air trench.

3. The intelligent design method for a wideband infrared polarization imaging device according to claim 1, characterized in that, The forward design network uses a loss function that combines normalized MSE and normalized Log-MSE.

4. The intelligent design method for a wideband infrared polarization imaging device according to claim 1, characterized in that, The convolutional module of the reverse-designed network includes a first convolutional layer, a first pooling layer, a second convolutional layer, and a second pooling layer connected in sequence.

5. The intelligent design method for a wideband infrared polarization imaging device according to claim 1, characterized in that, The reverse design network described above uses the MSE loss function.

6. A subwavelength metal grating, wherein the subwavelength metal grating is implemented based on the method according to any one of claims 1-5, characterized in that, The subwavelength metal grating adopts a bimetal-dielectric-metal structure, which uses air trenches to separate the bimetal-dielectric-metal. The metal is aluminum, the dielectric is silicon oxide, the substrate is aluminum oxide, and the wavelength range is 1~12μm.

Images