A wide-band laser protection imaging method and device

Abstract

The application discloses a wide-spectrum laser protection imaging method and device, and relates to the field of photoelectric imaging, which comprises the following steps: attaching a diffractive optical element in front of a camera lens to obtain a wavefront coding camera, modeling the wavefront coding camera, and determining a point spread function of the wavefront coding camera; constructing a two-stage image restoration network and training the two-stage image restoration network; collecting a coded defocus image through the wavefront coding camera, and processing the coded defocus image by using the trained two-stage image restoration network to obtain a predicted image. The application greatly reduces the risk that an imaging system is damaged by any laser in a wide-spectrum range, and the design of the two-stage image restoration network ensures high-quality imaging under laser interference.

Description

Technical Field

[0001] This invention relates to the field of optoelectronic imaging technology, and more specifically to a broadband laser protection imaging method and apparatus. Background Technology

[0002] Optoelectronic imaging systems (cameras) have become core tools for information acquisition, widely used in daily life and scientific research. Traditional cameras are typically designed with a highly focused point spread function (PSF) to achieve high-quality imaging. When the camera is exposed to a light beam (such as a laser), this structure generates a significant energy gain on the sensor surface. However, CCD / CMOS sensors are extremely sensitive to high-intensity light. When the peak light intensity focused on the sensor exceeds its damage threshold, it will cause hardware damage and lead to permanent loss of camera functionality. At the same time, the increasing prevalence of compact laser sources and the widespread application of lidar in everyday scenarios such as autonomous driving and drones further exacerbate the risk of camera damage.

[0003] Existing imaging technologies primarily employ spectral filtering, wavefront-coded laser protection based on specific phase functions, and lensless imaging techniques for laser protection. However, spectral filtering typically utilizes selective absorption, multilayer dielectric film reflection, and holographic phase grating diffraction to protect against lasers of specific wavelengths. This requires pre-determining the wavelength of the interfering laser, making it ineffective when the interfering laser is unknown. Wavefront-coded laser protection based on specific phase functions, by designing phase functions for specific wavelengths, can improve the laser protection capability of the imaging system. However, due to the dispersion effect of the phase element, the laser protection band is not wide enough. Furthermore, as laser protection capability increases, the quality of the recovered image rapidly declines, making it difficult to achieve a balance between protection capability and imaging quality. Lensless imaging systems, by placing a phase or amplitude modulation element in front of the sensor, can acquire images. However, due to the lack of a focusing lens, it is difficult to recover the scene with high quality from the acquired blurred image, resulting in poor image quality.

[0004] Therefore, how to provide an imaging method with broadband laser protection is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] In view of this, the present invention provides a broadband laser protection imaging method and apparatus, aiming to provide a solution for an optoelectronic imaging system that can maintain high-quality imaging while possessing laser protection capabilities. It not only maintains the high imaging quality of large-area optoelectronic imaging systems but also has broadband laser protection capabilities far superior to conventional optoelectronic imaging systems, effectively overcoming the shortcomings of existing optoelectronic imaging systems in laser protection and possessing broad application adaptability. To achieve the above objectives, the present invention adopts the following technical solution:

[0006] This invention discloses a broadband laser protection imaging method, comprising:

[0007] A wavefront-coded camera is obtained by attaching diffractive optical elements in front of the camera lens, and the wavefront-coded camera is modeled to determine the point spread function of the wavefront-coded camera.

[0008] Construct and train a two-stage image inpainting network;

[0009] The wavefront-coded camera acquires coded defocused images, and the trained two-stage image inpainting network is used to process the images to obtain predicted images.

[0010] Furthermore, the modeling of the wavefront-coded camera specifically involves: abstracting the wavefront-coded camera as a main lens with focal length f attached to a diffractive optical element, and a photoelectric sensor; assuming that the distance between the laser source and the wavefront-coded camera exceeds the Rayleigh range, and that both the laser and the incoherent information light are considered as plane waves, the point spread function in the Cartesian coordinate system is determined according to scalar diffraction theory.

[0011] Furthermore, the point spread function is expressed as:

[0012]

[0013] Where (s,t) represents the spatial coordinates at the camera pupil plane, and (x,y) represents the spatial coordinates at the camera sensor plane; The complex amplitude transmittance of the diffractive optical element and the lens is represented by λ, the incident light wavelength is represented by i, the imaginary unit is represented by k = 2π / λ, the wave number is represented by n, the refractive index of the diffractive optical element is represented by h(s,t), the surface height distribution of the diffractive optical element is represented by h(s,t), the natural constant is represented by e, the focal length of the lens is represented by f, and the aperture of the diffractive optical element is represented by A(s,t), that is, a circular binary aperture stop with a diameter of D.

[0014] Furthermore, the wavefront-coded camera operates in the spectral range [λ]. min ,λ max Within, the captured coded defocused image I c Represented as:

[0015]

[0016] Where S represents the truncation operation, I(λ) represents the ideal image, α represents the laser intensity, δ represents the Dirac function, (Δx,Δy) represents the random displacement of the laser glare on the sensor surface as a function of the incident angle, and p laser (λ) represents the energy distribution of the interfering laser with wavelength λ at the sensor plane, T c Represents the RGB spectral response curve, η~N(0,σ 2) represents Gaussian readout noise.

[0017] Furthermore, in the first stage of the two-stage image inpainting network, a U-Net architecture is adopted with an aggregation context transformation module as the core. Saturated pixels are filled with context inference prediction values ​​to remove laser glare from the encoded defocused image and generate a blurred image without laser interference.

[0018] In the second stage of the two-stage image inpainting network, the blurred image without laser interference and the point spread function are used as inputs, and the image is deblurred by Wiener filtering and deconvolution; then, residual U-Net is used for noise suppression and fine inpainting to obtain the predicted image.

[0019] Furthermore, the network structure of the first stage consists of three convolutional layers, eight aggregation context transformation modules, and three convolutional layers connected in sequence. The input of the first stage is the encoded defocused image and the binary mask of the saturated description region. After the input is downsampled by the first three convolutional layers, it is processed by the eight aggregation context transformation modules and then upsampled by the last three convolutional layers to generate the laser-free blurred image.

[0020] The aggregation context transformation module includes a convolutional fusion branch and a residual branch. The convolutional fusion branch includes four parallel convolutional modules and one convolutional fusion module. The four convolutional modules perform convolution and activation function processing on the input feature map, and then the output is fused by the convolutional fusion module. The residual branch convolves the input feature map and then performs a weighted residual connection with the output of the convolutional fusion branch to obtain the final output of the aggregation context transformation module.

[0021] Furthermore, the network structure of the second stage consists of a Wiener filter deconvolution module and a residual U-Net connected in sequence; taking the blurred image without laser interference and the point spread function as input, the image is deblurred by the Wiener filter deconvolution module to obtain the output feature map, which is then processed by the residual U-Net and concatenated with the output feature map of the Wiener filter deconvolution module. Finally, the predicted image is obtained after processing by the three residual modules.

[0022] The residual module includes a first convolutional layer, a first normalization layer, a first activation function layer, a second convolutional layer, a second normalization layer, and a second activation function layer connected in sequence. The feature map input to the residual module is subjected to convolution, normalization, activation, secondary convolution, and normalization operations, and then activated by the second activation function layer together with the feature map input to the residual module to obtain the final output of the residual module.

[0023] Furthermore, the overall loss function of the two-stage image inpainting network is expressed as:

[0024]

[0025] in, Indicates the total loss;

[0026] This indicates the glare reduction guidance loss term. A blurred image indicating no laser glare interference. Compared with the original blurred image I b L1 loss between; Indicates the saturated region with I b The L1 loss between them, M represents the binary mask describing the saturation region, and ⊙ represents the element-wise product operation; A blurred image indicating no laser glare interference. Compared with the original blurred image I b L1 loss between deep feature Gram matrices, φ i E represents the activation map of the i-th layer of the pre-trained feature extraction network VGG19. i α1 = 1, α2 = 1, α3 = 30 represent the mathematical expectation; α1 = 1, α2 = 1, α3 = 30 represent the weight coefficients of each loss function;

[0027] This represents the image deblurring guided loss term. The predicted image L1 loss between the original sharp image I and the original sharp image I; This indicates the L1 loss in the frequency domain. β1 = 1, β2 = 0.05 represent the weight coefficients of each loss function;

[0028] This indicates the laser protection guidance loss item. γ1=200λ target ={436nm,486nm,546nm,588nm,656nm};

[0029] This represents the point spread function constraint loss term, used to constrain the point spread function within a finite region. W(x,y) is a rectangular mask with a side length of 200 pixels. The value within the mask is 0, and the value outside the mask is 1.

[0030] This represents the constraint loss term for diffractive optical elements, used to limit the height h of the diffractive optical element within [0, h]. max Within the range, γ2=10 4 hmax =1.5μm,

[0031] The present invention also discloses a broadband laser protection imaging device, comprising: a diffractive optical element, an imaging lens, an area array detector, an image processor, and an output display device;

[0032] The diffractive optical element is attached to the main lens of the imaging lens. The diffractive optical element encodes the waveform of the incident light, and the incident light forms an coded defocused image after passing through the diffractive optical element and the imaging lens.

[0033] The area array detector captures the image signal of the encoded defocused image and sends it to the image processor;

[0034] The image processor uses a pre-trained two-stage image inpainting network to restore the coded defocused image to obtain a clear image, which is then displayed through the output display device.

[0035] As can be seen from the above technical solutions, compared with the prior art, the present invention discloses a broadband laser protection imaging method and device. By introducing phase modulation, the point spread function of the system is much larger than that of conventional imaging systems, thereby greatly reducing the peak energy of interfering lasers in the focal plane and greatly reducing the risk of the imaging system being damaged by any laser in the broadband range. At the same time, the design of the two-stage image restoration network ensures high-quality imaging under laser interference. Attached Figure Description

[0036] 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 embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0037] Figure 1 This is a schematic diagram of the overall process of the present invention.

[0038] Figure 2 This is a schematic diagram of the network structure provided by the present invention.

[0039] Figure 3 This is a schematic diagram of the device structure provided by the present invention.

[0040] Figure 4 This is a schematic diagram of the height distribution of the diffractive optical element provided by the present invention.

[0041] Figure 5 This is a schematic diagram of the experimental energy suppression ratio provided by the present invention.

[0042] Figure 6 This is a schematic diagram illustrating the broadband laser protection capability characterization provided by the present invention.

[0043] Figure 7 This is a schematic diagram of the experimental imaging results of the anti-interference laser camera provided by the present invention. Detailed Implementation

[0044] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0045] This invention discloses a broadband laser protection imaging method, comprising:

[0046] A wavefront-coded camera is obtained by attaching diffractive optical elements in front of the camera lens, and the wavefront-coded camera is modeled to determine the point spread function of the wavefront-coded camera.

[0047] Construct and train a two-stage image inpainting network;

[0048] Encoded defocused images are acquired using a wavefront-coded camera, and the predicted images are obtained by processing them using a trained two-stage image inpainting network.

[0049] Specifically, for the broadband laser protection targets in this invention, the height of each pixel of the diffractive optical element (DOE) is used as an optimization variable to maximize the degree of freedom in optical modulation. For example... Figure 1 As shown, to ensure high-quality imaging capabilities of the imaging system under laser interference, this invention proposes a two-stage image inpainting network. Guided by a weighted sum of multiple loss functions, a joint optimization framework of optics and algorithm is constructed to achieve optimal design of the DOE (Optical Design Object) and network weights. Finally, the DOE is fabricated using photolithography, and a prototype is built to realize the proposed optoelectronic imaging system. The broadband laser protection capability and imaging capability in outdoor scenes of the imaging system are verified by constructing optical paths.

[0050] In a specific embodiment, the wavefront-coded camera is modeled as follows: the wavefront-coded camera is abstracted as a main lens with focal length f attached to a diffractive optical element, and a photoelectric sensor; assuming that the distance between the laser source and the wavefront-coded camera exceeds the Rayleigh range, and that both the laser and the incoherent information light are regarded as plane waves, the midpoint spread function in the Cartesian coordinate system is determined according to the scalar diffraction theory.

[0051] Specifically, a typical laser hazard scenario can be described as follows: when a camera is focused on a far-field scene, an interfering laser with wavelength λ is incident on its aperture. The optical model of a wavefront-coded (WFC) camera can be simplified to a primary lens with focal length f, a diffractive optical element (DOE) attached, and a photoelectric sensor. It is assumed that the laser source is located beyond the Rayleigh range, and that both the laser and the incoherent information light can be considered as plane wavefronts. Therefore, the laser intensity distribution on the sensor surface is equivalent to the camera's point spread function (PSF). According to scalar diffraction theory, the point spread function in Cartesian coordinates is expressed as:

[0052]

[0053] Where (s,t) represents the spatial coordinates at the camera pupil plane, and (x,y) represents the spatial coordinates at the camera sensor plane; The complex amplitude transmittance of the diffractive optical element and the lens is represented by λ, the incident light wavelength is represented by i, the imaginary unit is represented by k = 2π / λ, the wave number is represented by n, the refractive index of the diffractive optical element is represented by h(s,t), the surface height distribution of the diffractive optical element is represented by h(s,t), the natural constant is represented by e, the focal length of the lens is represented by f, and the aperture of the diffractive optical element is represented by A(s,t), that is, a circular binary aperture stop with a diameter of D.

[0054] In one specific embodiment, in addition to laser glare, the point spread function also affects the camera's imaging process. Wavefront-coded cameras operate within the spectral range [λ...]. min ,λ max Within, the captured coded defocused image I c Represented as:

[0055]

[0056] Where S represents the truncation operation, I(λ) represents the ideal image, α represents the laser intensity, δ represents the Dirac function, (Δx,Δy) represents the random displacement of the laser glare on the sensor surface as a function of the incident angle, and p laser (λ) represents the energy distribution of the interfering laser with wavelength λ at the sensor plane, T c Represents the RGB spectral response curve, η~N(0,σ 2 ) represents Gaussian readout noise.

[0057] In one specific embodiment, the first stage of the two-stage image inpainting network adopts the U-Net architecture with the aggregation context transformation module as the core. It fills saturated pixels with context inference prediction values, removes laser glare from the encoded defocused image, and generates a blurred image without laser interference.

[0058] In the second stage of the two-stage image inpainting network, a blurred image without laser interference and a point spread function are used as inputs. The image is deblurred by Wiener filtering and deconvolution. Then, residual U-Net is used for noise suppression and fine inpainting to obtain the predicted image.

[0059] In one specific embodiment, the network structure of the first stage consists of three convolutional layers, eight aggregation context transformation modules, and three convolutional layers connected in sequence. The input of the first stage is an encoded defocused image and a binary mask describing the saturated region. After the input is downsampled by the first three convolutional layers, it is processed by the eight aggregation context transformation modules and then upsampled by the last three convolutional layers to generate a blurred image without laser interference.

[0060] The aggregation context transformation module includes a convolutional fusion branch and a residual branch. The convolutional fusion branch includes four parallel convolutional modules and one convolutional fusion module. The four convolutional modules perform convolution and activation function processing on the input feature map respectively, and then the output is fused by the convolutional fusion module. The residual branch convolves the input feature map and then performs a weighted residual connection with the output of the convolutional fusion branch to obtain the final output of the aggregation context transformation module.

[0061] In a specific embodiment, the network structure of the second stage consists of a Wiener filter deconvolution module and a residual U-Net connected in sequence. The image is deblurred by the Wiener filter deconvolution module and a point spread function as input. The output feature map is obtained by the image deblurring process of the Wiener filter deconvolution module. After processing by the residual U-Net, it is superimposed with the output feature map of the Wiener filter deconvolution module. Finally, the predicted image is obtained after processing by the three residual modules.

[0062] The residual module comprises a first convolutional layer, a first normalization layer, a first activation function layer, a second convolutional layer, a second normalization layer, and a second activation function layer connected in sequence. The feature map input to the residual module undergoes convolution, normalization, activation, secondary convolution, and normalization operations, and is then superimposed with the feature map input to the residual module through the second activation function layer and activated to obtain the final output of the residual module.

[0063] Specifically, to restore complex degraded images, this invention constructs a hybrid neural network that includes repair and decoding modules, such as... Figure 2As shown in part a of the diagram (where Convolution represents convolution, AOT Block represents the Aggregate Context Transformation module, Wiener filter represents the Wiener filter deconvolution module, and Residual Block represents the residual module), the corresponding neural network post-processing is divided into two stages: first, laser glare is removed, and then the image is deblurred. Since the intensity of the focused laser light on the sensor surface is usually much higher than the background light, interfering pixels are easily saturated, leading to information loss. In the first stage, the Aggregate Context Transformation module (AOT Block, as shown in part a) is used. Figure 2 As shown in section b), the core inpainting unit of the neural network fills saturated pixels with predicted values ​​through context inference. This neural network adopts a U-Net architecture, using a binary mask of the degraded image and the saturated description region as dual inputs. After downsampling, the input is processed through eight AOT modules and then upsampled to generate a blurred image without laser interference. Subsequently, the blurred image without laser interference generated in stage one is input into stage two for deblurring. To balance restoration quality and processing efficiency, a physical model-assisted neural network was developed: first, image deblurring is achieved through Wiener filtering deconvolution, and then noise suppression and fine inpainting are performed using Residual U-Net (Res-Unet). The structure of the residual module is as follows: Figure 2 As shown in part c.

[0064] In a specific embodiment, to achieve broadband laser protection and high-quality imaging, the overall loss function of the two-stage image inpainting network is expressed as:

[0065]

[0066] Where L represents the total loss;

[0067] L g This indicates the glare reduction guidance loss term. A blurred image indicating no laser glare interference. Compared with the original blurred image I b L1 loss between; Indicates the saturated region with I b The L1 loss between them, M represents the binary mask describing the saturation region, and ⊙ represents the element-wise product operation; A blurred image indicating no laser glare interference. Compared with the original blurred image I b L1 loss between deep feature Gram matrices, φ i This represents the activation map of the i-th layer of the pre-trained feature extraction network VGG19. α1 = 1, α2 = 1, α3 = 30 represent the mathematical expectation; α1 = 1, α2 = 1, α3 = 30 represent the weight coefficients of each loss function;

[0068] This represents the image deblurring guided loss term. Represents the predicted image L1 loss between the original sharp image I and the original sharp image I; This represents the L1 loss in the frequency domain, F represents the Fourier transform operator; β1=1, β2=0.05 represent the weight coefficients of each loss function;

[0069] This indicates the laser protection guidance loss item. λ target ={436nm,486nm,546nm,588nm,656nm};

[0070] This represents the point spread function constraint loss term, used to constrain the point spread function within a finite region. W(x,y) is a rectangular mask with a side length of 200 pixels. The value within the mask is 0, and the value outside the mask is 1.

[0071] This represents the constraint loss term for diffractive optical elements, used to limit the height h of the diffractive optical element within [0, h]. max Within the range, γ2=10 4 h max =1.5μm,

[0072] In one specific embodiment, the specific design parameters of the present invention are as follows: focal length f = 50mm, aperture diameter D = 4mm; sensor pixel size is 4.8×4.8μm, resolution is 572×572; camera operating wavelength is discretized into three bands: 436nm, 546nm, and 656nm. The training dataset (512×512) is derived from the DIV2K dataset, containing 1324 training images and 50 test images; the Adam optimizer is used for 150 rounds of training, with a fixed batch size of 1; the initial learning rate of the diffractive optical element (DOE) is 10. -9 The initial learning rate of the neural network is 10. -4 The learning rate decays by 0.8 times every 50 rounds; after the first 50 rounds of training, the DOE is highly quantized into 8 discrete levels for processing. Training on an NVIDIA GeForce RTX 3080Ti single graphics card takes approximately three days.

[0073] Figure 4The height distribution map of the designed DOE is shown. The sample was fabricated using photolithography, with a chromium-plated layer covering the non-functional areas. The DOE was mounted on the front group of a composite lens and integrated with an imaging sensor to finally construct a prototype laser-resistant camera system. To verify the laser suppression effect of the laser-resistant camera, the intensity distribution of the laser spot on the sensor surface of both the conventional camera and the laser-resistant camera was measured in the optical path. The beam emitted from the supercontinuum laser (NKT Ph otonics SuperK COMPACT, 450-2400nm) was first modulated by a filter to generate quasi-monochromatic light of discrete wavelengths; after passing through a spatial filter (including objective lens and pinhole assembly), it was collimated into a plane wave by an achromatic doublet lens 1 (Lbtek MAD410-A, f=150mm); the beam was captured by both the laser-resistant camera and the conventional camera, and their respective point spread functions (PSF) were obtained. In the quantitative evaluation, the energy suppression ratio (ESR) was calculated by measuring the peak single-pixel power of the normalized PSF (energy unit).

[0074]

[0075] Among them, PSF anti-laser The point spread function (PSF) represents the laser-resistant camera. conv This refers to the PSF of a traditional camera. Figure 5 The experimental energy suppression ratio (ESR) (red data points) in the 473-688 nm wavelength range is shown, and the simulated ESR results for 100% modulation efficiency (μ) (red dashed line) and 99% modulation efficiency (blue dashed line) are compared. It is clearly observed that the experimental results are in high agreement with the simulated predictions. Within the 215 nm bandwidth of the visible spectrum, the ESR exceeds -40 dB, achieving broadband laser protection. It is worth noting that due to the imperfect modulation efficiency of the DOE, the actual ESR is slightly lower than the ideal predicted value of 100% modulation efficiency.

[0076] Figure 6 The intensity distribution of the designed PSF at six wavelengths is visualized. The PSFs obtained from experiments and simulations show high consistency across all wavelengths. The designed PSF (width > 50 pixels) has a significantly larger spatial size compared to the traditional Airy disk PSF (full width at half maximum (FWHM) < 3 pixels), thus effectively dispersing the focused laser energy to reduce the sensor peak intensity (ESR > -40 dB). Figure 7 Experimental imaging results of an anti-jamming laser camera using the network decoding proposed in this invention are shown, such as... Figure 7 As shown in part a, the fine structure of the laser-free scene is effectively reconstructed, and color fidelity is maintained, indicating that the camera possesses high-quality color imaging capabilities; as Figure 7 As shown in part b, although the oversaturated areas of the laser scene may exhibit slight distortion, the scene outline remains recognizable thanks to the algorithm reconstruction.

[0077] This invention also discloses a broadband laser protection imaging device, such as... Figure 3 As shown, it includes: diffractive optical elements, imaging lens (lens in the figure), area array detector (camera sensor in the figure), image processor (computer in the figure), and output display device;

[0078] A diffractive optical element is attached to the main lens of the imaging lens. The diffractive optical element encodes the waveform of the incident light, and the incident light forms an coded defocused image after passing through the diffractive optical element and the imaging lens.

[0079] An area array detector captures image signals that encode defocused images and sends them to an image processor;

[0080] An image processor uses a pre-trained two-stage image inpainting network to restore a coded defocused image to obtain a clear image, which is then displayed through an output display device. The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.

[0081] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A broadband laser protection imaging method, characterized in that, include: A wavefront-coded camera is obtained by attaching diffractive optical elements in front of the camera lens, and the wavefront-coded camera is modeled to determine the point spread function of the wavefront-coded camera. Construct and train a two-stage image inpainting network; The wavefront-coded camera acquires coded defocused images, and the trained two-stage image inpainting network is used to process the images to obtain predicted images.

2. The broadband laser protection imaging method according to claim 1, characterized in that, The modeling of the wavefront-coded camera specifically involves: abstracting the wavefront-coded camera as a main lens with focal length f attached to a diffractive optical element, and a photoelectric sensor; assuming that the distance between the laser source and the wavefront-coded camera exceeds the Rayleigh range, both the laser and the incoherent information light can be regarded as plane waves, and the point spread function in the Cartesian coordinate system is determined according to the scalar diffraction theory.

3. The broadband laser protection imaging method according to claim 1, characterized in that, The point spread function is expressed as: Where (s,t) represents the spatial coordinates at the camera pupil plane, and (x,y) represents the spatial coordinates at the camera sensor plane; The complex amplitude transmittance of the diffractive optical element and the lens is represented by λ, the incident light wavelength is represented by i, the imaginary unit is represented by k = 2π / λ, the wave number is represented by n, the refractive index of the diffractive optical element is represented by h(s,t), the surface height distribution of the diffractive optical element is represented by h(s,t), the natural constant is represented by e, the focal length of the lens is represented by f, and the aperture of the diffractive optical element is represented by A(s,t), that is, a circular binary aperture stop with a diameter of D.

4. The broadband laser protection imaging method according to claim 3, characterized in that, The wavefront-coded camera operates in the spectral range [λ]. min ,λ max Within, the captured coded defocused image I c Represented as: Where S represents the truncation operation, I(λ) represents the ideal image, α represents the laser intensity, δ represents the Dirac function, (Δx,Δy) represents the random displacement of the laser glare on the sensor surface as a function of the incident angle, and p laser (λ) represents the energy distribution of the interfering laser with wavelength λ at the sensor plane, T c Represents the RGB spectral response curve, η~N(0,σ 2 ) represents Gaussian readout noise.

5. The broadband laser protection imaging method according to claim 1, characterized in that, In the first stage of the two-stage image inpainting network, a U-Net architecture is adopted with an aggregation context transformation module as the core. Saturated pixels are filled with context inference prediction values ​​to remove laser glare from the encoded defocused image and generate a blurred image without laser interference. In the second stage of the two-stage image inpainting network, the blurred image without laser interference and the point spread function are used as inputs, and the image is deblurred by Wiener filtering and deconvolution; then, residual U-Net is used for noise suppression and fine inpainting to obtain the predicted image.

6. The broadband laser protection imaging method according to claim 5, characterized in that, The network structure of the first stage consists of three convolutional layers, eight aggregation context transformation modules, and three convolutional layers connected in sequence. The input of the first stage is the encoded defocused image and a binary mask describing the saturated region. After the input is downsampled by the first three convolutional layers, it is processed by the eight aggregation context transformation modules and then upsampled by the last three convolutional layers to generate the blurred image without laser interference. The aggregation context transformation module includes a convolutional fusion branch and a residual branch. The convolutional fusion branch includes four parallel convolutional modules and one convolutional fusion module. The four convolutional modules perform convolution and activation function processing on the input feature map, and then the output is fused by the convolutional fusion module. The residual branch convolves the input feature map and then performs a weighted residual connection with the output of the convolutional fusion branch to obtain the final output of the aggregation context transformation module.

7. The broadband laser protection imaging method according to claim 5, characterized in that, The network structure of the second stage consists of a Wiener filter deconvolution module and a residual U-Net connected in sequence. Taking the blurred image without laser interference and the point spread function as input, the image is deblurred by the Wiener filter deconvolution module to obtain the output feature map. After being processed by the residual U-Net, it is superimposed with the output feature map of the Wiener filter deconvolution module. Finally, the predicted image is obtained after being processed by three residual modules. The residual module includes a first convolutional layer, a first normalization layer, a first activation function layer, a second convolutional layer, a second normalization layer, and a second activation function layer connected in sequence. The feature map input to the residual module is subjected to convolution, normalization, activation, double convolution, and normalization operations, and then superimposed with the feature map input to the residual module and activated by the second activation function layer to obtain the final output of the residual module.

8. The broadband laser protection imaging method according to claim 4, characterized in that, The overall loss function of the two-stage image inpainting network is expressed as: in, Indicates the total loss; This indicates the glare reduction guidance loss term. A blurred image indicating no laser glare interference. Compared with the original blurred image I b L1 loss between; Indicates the saturated region with I b The L1 loss between them, M represents the binary mask describing the saturation region, and ⊙ represents the element-wise product operation; A blurred image indicating no laser glare interference. Compared with the original blurred image I b L1 loss between deep feature Gram matrices, φ i This represents the activation map of the i-th layer of the pre-trained feature extraction network VGG19. α1 = 1, α2 = 1, α3 = 30 represent the mathematical expectation; α1 = 1, α2 = 1, α3 = 30 represent the weight coefficients of each loss function; This represents the image deblurring guided loss term. The predicted image L1 loss between the original sharp image I and the original sharp image I; This indicates the L1 loss in the frequency domain. β1 = 1, β2 = 0.05 represent the weight coefficients of each loss function; This indicates the laser protection guidance loss item. γ1=200λ target ={436nm,486nm,546nm,588nm,656nm}; This represents the point spread function constraint loss term, used to constrain the point spread function within a finite region. W(x,y) is a rectangular mask with a side length of 200 pixels. The value within the mask is 0, and the value outside the mask is 1. This represents the constraint loss term for diffractive optical elements, used to limit the height h of the diffractive optical element within [0, h]. max Within the range, γ2=10 4 h max =1.5μm, 9. A broadband laser protection imaging device, employing the broadband laser protection imaging method according to any one of claims 1 to 8, characterized in that, include: Diffractive optical elements, imaging lenses, area array detectors, image processors, and output display devices; The diffractive optical element is attached to the main lens of the imaging lens. The diffractive optical element encodes the waveform of the incident light, and the incident light forms an coded defocused image after passing through the diffractive optical element and the imaging lens. The area array detector captures the image signal of the encoded defocused image and sends it to the image processor; The image processor uses a pre-trained two-stage image inpainting network to restore the coded defocused image to obtain a clear image, which is then displayed through the output display device.

Images