Reflective optical system based on computational imaging driven machining simplification

Abstract

The application relates to a reflection optical system based on calculation imaging driving processing simplification, which comprises a reflection optical module, a detector and an image reconstruction system, the reflection optical module comprises a first free curved mirror, a second even non-spherical mirror and a third free curved mirror, the detector is coincident with the focal plane of the reflection optical module, the image reconstruction system comprises an NPU image data processor, and the NPU image data processor runs an image restoration algorithm model; incident light is sequentially focused on the detector through the first free curved mirror, the second even non-spherical mirror and the third free curved mirror, the detection result is transmitted to the image reconstruction system for processing through the detector, and an image with image quality close to the diffraction limit is output. Compared with the prior art, the application can reduce the optical system processing requirement by at least one order of magnitude, and has the imaging quality close to the diffraction limit.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of reflective optical system structure, and in particular to a reflective optical system based on calculation imaging driving processing simplification. BACKGROUND

[0002] The reflective optical system structure is the mainstream configuration of space optical load, such as the invention disclosed in the publication No. CN115453732A, which comprises a primary mirror, a secondary mirror, a first turning mirror, a second turning mirror, a third mirror and a focal plane detector assembly. The optical axes of the above-mentioned turning mirrors, mirrors and focal plane detector assembly are located in the same plane. The aperture stop of the entire optical system coincides with the primary mirror. The light from the scene passes through the reflection of the primary mirror and the secondary mirror in turn, reaches the first turning mirror, the first turning mirror deflects the light to the second turning mirror, the light is deflected to the third mirror through the second turning mirror, and finally reaches the focal plane detector assembly through the reflection of the third mirror.

[0003] In order to realize high-quality imaging, the surface shape of the reflective optical system is changed from the traditional spherical surface to the aspherical surface and the free-form surface, and the nanometer-level processing precision is required, and the root mean square (RMS) value of the processing surface error is usually better than 1 / 50λ. In order to realize such high processing precision, the current processing approaches are: single-point diamond turning based on compensation correction, small grinding head precision modification, ion beam ultra-precision modification and magnetorheological ultra-precision modification. Although these processing methods can achieve nanometer-level processing precision, they consume a large amount of time and are expensive, which cannot meet the requirements of commercial aerospace rapid manufacturing, batch production and modular production. SUMMARY

[0004] The purpose of the present application is to overcome the defects of the prior art that the reflective optical system requires nanometer-level processing precision, consumes a large amount of time and is expensive, and to provide a reflective optical system based on calculation imaging driving processing simplification.

[0005] The purpose of the present application can be achieved by the following technical solutions:

[0006] A reflective optical system based on calculation imaging driving processing simplification, comprising a reflective optical module, a detector and an image reconstruction system, wherein the reflective optical module comprises a first free-form mirror, a second even aspherical mirror and a third free-form mirror, the detector coincides with the focal plane of the reflective optical module, and the image reconstruction system comprises an NPU image data processor running an image restoration algorithm model.

[0007] The incident light is sequentially focused by the first free-form mirror, the second even-order aspheric mirror and the third free-form mirror and then is incident on the detector, and the detection result is transmitted to the image reconstruction system for processing, and an image with image quality close to the diffraction limit is output.

[0008] Further, the F number of the reflective optical module is within a range of 1-14, and the entrance pupil diameter is within a range of 15mm-600mm.

[0009] Further, the machining surface shape of the first free-form mirror, the second even-order aspheric mirror and the third free-form mirror all satisfy that the root mean square is better than λ / 2, wherein λ is the wavelength of light.

[0010] Further, the detector is a visible light COMS detector.

[0011] Further, the image restoration algorithm model is a deep learning model based on physical prior, the physical prior is a Wiener filtering algorithm model based on a point spread function of the reflective optical system, and the deep learning model is a dynamic convolution deblurring model based on a field of view attention mechanism.

[0012] Further, the Wiener filtering algorithm model based on the point spread function of the reflective optical system performs filtering processing on an input picture and a corresponding point spread function PSF, and adds a regularization parameter, and finally obtains a deconvolution image.

[0013] Further, the input of the dynamic convolution deblurring model based on the field of view attention mechanism includes a field of view coordinate matrix and a deconvolution image.

[0014] The dynamic convolution deblurring model based on the field of view attention mechanism includes an FOV encoder and a deformable decoder connected with each other.

[0015] Further, the FOV encoder includes FOV modules, a first encoding module, a second encoding module, a third encoding module and a fourth encoding module connected in sequence, the first encoding module includes convolution layers Conv and ResBlock connected in sequence, and the second encoding module, the third encoding module and the fourth encoding module each include convolution layers Conv with a stride of 2 and ResBlock connected in sequence.

[0016] The deformable decoder includes a first decoding module, a second decoding module, a third decoding module, a fourth decoding module, a KPN module, a first connection module, a second connection module, a third connection module and a fourth connection module.

[0017] The first decoding module includes a deformable ResBlock; the second and third decoding modules each include a deconvolutional layer Deconv, a convolutional layer Conv, and a deformable ResBlock connected in sequence; the fourth decoding module includes a deconvolutional layer Deconv, a convolutional layer Conv, a deformable ResBlock, and a convolutional layer Conv connected in sequence; the first connecting module includes an offset adjustment layer Offset Transfer; the second, third, and fourth connecting modules each include a bilinear upsampling layer Bilinear Upsampling and an offset adjustment layer OffsetTransfer connected in sequence.

[0018] The fourth encoding module, the first decoding module, the second decoding module, the third decoding module, the fourth decoding module, and the KPN module are connected in sequence. The input of the first connecting module is connected to the ResBlock of the fourth encoding module, and the output is used as the input of the first decoding module.

[0019] In the second connection module, the input of the bilinear upsampling layer is the output of the first connection module, the input of the offset transfer layer also includes the output of the convolutional layer Conv of the second decoding module, the output of the offset transfer layer is connected to the deformable ResBlock of the second decoding module; and the input of the convolutional layer Conv of the second decoding module also includes the output of the ResBlock of the third encoding module.

[0020] The input of the bilinear upsampling layer in the third connecting module is the output of the offset transfer layer of the second connecting module. The input of the offset transfer layer also includes the output of the convolutional layer Conv of the third decoding module. The output of the offset transfer layer is connected to the deformable ResBlock of the third decoding module. Furthermore, the input of the convolutional layer Conv of the third decoding module also includes the output of the ResBlock of the second encoding module.

[0021] The input of the Bilinear Upsampling layer in the fourth connecting module is the output of the Offset Transfer layer of the third connecting module and the output of the first convolutional layer Conv of the fourth decoding module. The output of the Offset Transfer layer is connected to the deformable ResBlock of the fourth decoding module; and the input of the convolutional layer Conv of the fourth decoding module also includes the output of the ResBlock of the first encoding module.

[0022] The input of the KPN module is also connected to the input deconvolution image, and after being superimposed with the input feature map, the images are summarized and processed through a convolution operation to obtain the final output.

[0023] Furthermore, the FOV module first extracts features from the input field of view coordinate matrix FOV through convolutional layers and adds a sigmoid activation function to obtain the FOV feature map; then it processes the input deconvolutional image through convolutional layers to obtain the deconvolutional feature map; then it multiplies the FOV feature map and the deconvolutional feature map and superimposes them to obtain the output.

[0024] The deformable ResBlock extracts features based on the offset output by the corresponding connecting module and the input convolutional layer Conv signal. After convolution and superposition with the input convolutional layer Conv signal, the output is obtained.

[0025] Furthermore, the NPU image data processor is an AI chip with the capability of deploying neural networks at the edge, used to deploy image restoration algorithm models and meet the computing power required for video-level real-time inference.

[0026] Compared with the prior art, the present invention has the following advantages:

[0027] (1) This invention proposes to shift the focus of the reflective optical system from the system processing accuracy to computational imaging technology. An NPU image data processor is set in the image reconstruction system, and an image restoration algorithm model is deployed to process the acquired image data in real time, thereby realizing efficient image restoration, recognition and analysis, thus supporting the system to achieve accurate imaging and high-quality image reconstruction under complex optical conditions. This can effectively reduce the processing requirements of optical components. The processing accuracy requirements are reduced by more than an order of magnitude compared with traditional systems, which can greatly improve the processing efficiency of optical components, reduce the cost of optical systems, and improve imaging quality.

[0028] (2) The present invention introduces computational imaging for aberration correction and uses appropriate image processing algorithms to effectively restore the degraded image of the optical system to an excellent state.

[0029] (3) The physical prior deep learning model of this invention organically combines the Wiener filtering algorithm based on the system point spread function and the dynamic convolution deblurring algorithm based on the field attention mechanism, which can achieve high-quality restoration of degraded images when the spatial degradation changes drastically and there are large and small aberrations. The restored image performance is close to the diffraction limit. Attached Figure Description

[0030] Figure 1This is a schematic diagram of a reflective optical system based on computational imaging-driven processing simplification provided in an embodiment of the present invention;

[0031] Figure 2 This is an off-axis three-reflector path diagram provided in an embodiment of the present invention;

[0032] Figure 3 This is a schematic diagram of the processing error boundary of an optical element provided in an embodiment of the present invention;

[0033] Figure 4 This is a schematic diagram of an image restoration algorithm model provided in an embodiment of the present invention;

[0034] Figure 5 This is a schematic diagram of an FOV_KPN neural network model provided in an embodiment of the present invention;

[0035] Figure 6 This is a comparison image before and after image restoration provided in an embodiment of the present invention;

[0036] In the figure, 1 is a reflective optical module, 101 is a first freeform surface mirror, 102 is a second even-order aspherical mirror, 103 is a third freeform surface mirror, 2 is a detector, 3 is an image reconstruction system, 301 is an image restoration algorithm model, and 302 is an NPU image data processor. Detailed Implementation

[0037] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0038] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0039] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0040] In the description of this invention, it should be noted that the terms "center," "upper," "lower," "left," "right," "vertical," "horizontal," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings, or the orientation or positional relationship in which the product of this invention is usually placed during use. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0041] It should be noted that the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0042] Furthermore, terms such as "horizontal" and "vertical" do not imply that components must be absolutely horizontal or suspended, but rather that they can be slightly tilted. For example, "horizontal" simply means that its direction is more horizontal than "vertical," and does not mean that the structure must be completely horizontal, but can be slightly tilted.

[0043] Example 1

[0044] To address the high technical requirements and cost associated with the reflective surface design of reflective optical systems in diamond machining, computational imaging can be used to shift the focus of aberration correction from optical hardware to image processing algorithms. This embodiment proposes a simplified reflective optical system driven by computational imaging. By introducing computational imaging for aberration correction, the required machining accuracy is reduced while maintaining image quality.

[0045] like Figure 1 As shown, it specifically includes a reflective optical module 1, a detector 2, and an image reconstruction system 3. The reflective optical module 1 includes a first freeform surface mirror 101, a second even-order aspherical mirror 102, and a third freeform surface mirror 103. The detector 2 coincides with the focal plane of the reflective optical module 1. The image reconstruction system 3 includes an NPU image data processor 302, which runs an image restoration algorithm model 301.

[0046] The incident light passes through the first freeform surface mirror 101, the second even-order aspherical mirror 102 and the third freeform surface mirror 103 in sequence and is then focused onto the detector 2. The detector 2 transmits the detection results to the image reconstruction system 3 for processing and outputs an image with near-diffraction-limited image quality.

[0047] In this embodiment, the reflective optical module employs an off-axis three-mirror system, such as... Figure 2 As shown. Without adding processing error data, the system has a field of view of 8.0° × 6.4°, an F number of 2.8, a wavelength range covering 400 nm to 1050 nm, a maximum RMS spot diameter of 0.988 μm across the entire field of view, and an MTF value exceeding 0.78 at a Nyquist frequency of 72 lp / mm.

[0048] In reflective optical modules, the smaller the F-number, the larger the entrance pupil diameter, the stronger the light-gathering ability, the larger the effective entrance pupil, and the smaller the corresponding field of view. Based on the detector model and parameters, and by compromising between the field of view and the light-gathering ability, the actual parameters should be selected. Therefore, the F-number of the reflective optical module should be limited to 1~14, and the entrance pupil diameter should be limited to 15mm~600mm.

[0049] In reflective optical module structures, the processing precision of the reflective surface determines the imaging quality. Smaller processing surface errors result in better imaging quality, and vice versa. Generally, to obtain near-diffraction-limited imaging quality, an optical system requires a processing error of less than λ / 50. This invention utilizes computational imaging technology to reduce the requirements for reflective surface processing quality. Based on preliminary experiments and image restoration experience, an MTF greater than 0.3 in the imaging system results in relatively good image restoration. Therefore, the processing error level corresponding to ensuring a system MTF greater than 0.3 was calculated. The final calculated processing error boundary is as follows: Figure 3 As shown, in the system with the added image preprocessing deconvolution module, the processing errors of the first freeform surface mirror 101, the second even-order aspherical mirror 102, and the third freeform surface mirror 103 should respectively satisfy the following requirements: primary mirror less than 0.675λ, secondary mirror less than 0.924λ, and third mirror less than 0.556λ. Therefore, it is determined that as long as the processing error of each surface shape of the optical system satisfies the requirement of having a root mean square (RMS) better than λ / 2, the requirements can be met.

[0050] In the detector image acquisition system 2, the actual detector selected is a visible light CMOS detector with a pixel size of 6.9μm and a total number of pixels of 1224*10244.

[0051] After the optical system is set up for PSF acquisition, the image is imported into the image reconstruction system for image restoration. The network structure is as follows: Figure 4 As shown.

[0052] The image reconstruction system 3 includes an NPU image data processor 302, which is an AI chip with the capability of edge deployment of neural networks. It has the computing power to deploy a deep learning-based image restoration algorithm model 301 and meet the real-time inference requirements for video-level 25Hz. This AI chip achieves efficient image restoration, recognition, and analysis by processing the acquired image data in real time, thereby supporting the system's accurate imaging and high-quality image reconstruction under complex optical conditions.

[0053] Image restoration algorithm model 301 is a deep learning model based on physical priors. The physical priors are Wiener filtering algorithm models based on the point spread function of a reflective optical system, and the deep learning model is a dynamic convolution deblurring model based on the field attention mechanism.

[0054] Network structure such as Figure 4 As shown, the Wiener filtering algorithm model based on the point spread function of the reflective optical system performs filtering processing on the input image and the corresponding point spread function PSFs, and adds regularization parameters to finally obtain the deconvolution image;

[0055] like Figure 5 As shown, the dynamic convolutional deblurring model based on the field of view attention mechanism, namely the FOV_KPN neural network, takes a deconvolutional image and a field of view coordinate matrix as input. Its network structure includes an interconnected FOV encoder and a deformable decoder. The FOV encoder includes a sequentially connected FOV module, a first encoding module, a second encoding module, a third encoding module, and a fourth encoding module. The first encoding module includes a sequentially connected convolutional layer Conv and a ResBlock. The second, third, and fourth encoding modules each include a sequentially connected convolutional layer Conv and a ResBlock with a stride of 2.

[0056] The deformable decoder includes a first decoding module, a second decoding module, a third decoding module, a fourth decoding module, a KPN module, a first connection module, a second connection module, a third connection module, and a fourth connection module;

[0057] The first decoding module includes a deformable ResBlock; the second and third decoding modules each include a deconvolutional layer (Deconv), a convolutional layer (Conv), and a deformable ResBlock connected in sequence; the fourth decoding module includes a deconvolutional layer (Deconv), a convolutional layer (Conv), a deformable ResBlock, and a convolutional layer (Conv) connected in sequence; the first connecting module includes an offset transfer layer; the second, third, and fourth connecting modules each include a bilinear upsampling layer (Bilinear Upsampling) and an offset transfer layer (Offset Transfer) connected in sequence.

[0058] The fourth encoding module, the first decoding module, the second decoding module, the third decoding module, the fourth decoding module, and the KPN module are connected in sequence. The input of the first connecting module is connected to the ResBlock of the fourth encoding module, and the output is used as the input of the first decoding module.

[0059] The input of the bilinear upsampling layer in the second connection module is the output of the first connection module; the input of the offset transfer layer also includes the output of the convolutional layer Conv of the second decoding module; the output of the offset transfer layer is connected to the deformable ResBlock of the second decoding module; and the input of the convolutional layer Conv of the second decoding module also includes the output of the ResBlock of the third encoding module.

[0060] The input of the bilinear upsampling layer in the third connecting module is the output of the offset transfer layer in the second connecting module. The input of the offset transfer layer also includes the output of the convolutional layer Conv in the third decoding module. The output of the offset transfer layer is connected to the deformable ResBlock in the third decoding module. Furthermore, the input of the convolutional layer Conv in the third decoding module also includes the output of the ResBlock in the second encoding module.

[0061] The input of the Bilinear Upsampling layer in the fourth connection module is the output of the Offset Transfer layer of the third connection module and the output of the first convolutional layer Conv of the fourth decoding module. The output of the Offset Transfer layer is connected to the deformable ResBlock of the fourth decoding module; and the input of the convolutional layer Conv of the fourth decoding module also includes the output of the ResBlock of the first encoding module.

[0062] The input of the KPN module is also connected to the input deconvolution image, and after being superimposed on the input feature map, the images are summarized and then processed through a convolution operation to obtain the final output.

[0063] The FOV module first extracts features from the input field of view (FOV) coordinate matrix through convolutional layers and adds a sigmoid activation function to obtain the FOV feature map; then it processes the input deconvolutional image through convolutional layers to obtain the deconvolutional feature map; finally, it multiplies the FOV feature map and the deconvolutional feature map and then superimposes them to obtain the output.

[0064] The deformable ResBlock extracts features based on the offset output from the corresponding connecting module and the input convolutional layer Conv signal. After convolution and superposition with the input convolutional layer Conv signal, the output is obtained.

[0065] The physical prior deep learning model organically combines the Wiener filtering algorithm based on the system point spread function and the dynamic convolution deblurring algorithm based on the field attention mechanism. It can achieve high-quality restoration of degraded images with drastic spatial degradation and large and small aberrations, and the restored image performance is close to the diffraction limit.

[0066] Based on the measured PSF of the optical system, this embodiment trains the network of the image restoration algorithm model 301 and performs real-world measurements on the trained system. The restoration results are as follows: Figure 6 As shown. Figure 6 Image (1) shows a comparison of the MTF of the images before and after recovery using the blade method test. Figure 6 Figure (2) shows a comparison between the actual captured image, the restored actual captured image, and the image captured by a Canon lens with good image quality. It can be seen that this method can restore the image to an MTF greater than 0.5, demonstrating superior performance in actual shooting.

[0067] In summary, the all-aluminum reflective optical system based on computational imaging-driven simplified processing provided in this embodiment reduces the processing accuracy requirement by more than an order of magnitude compared to traditional systems by introducing computational imaging for aberration correction. It also results in lower manufacturing costs, shorter development cycles, and better imaging quality.

[0068] The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.

Claims

1. A reflective optical system based on computational imaging-driven simplified processing, characterized in that, The system includes a reflective optical module (1), a detector (2), and an image reconstruction system (3). The reflective optical module (1) includes a first freeform surface mirror (101), a second even-order aspherical mirror (102), and a third freeform surface mirror (103). The detector (2) is aligned with the focal plane of the reflective optical module (1). The image reconstruction system (3) includes an NPU image data processor (302) that runs an image restoration algorithm model (301). The incident light is focused by the first freeform surface mirror (101), the second even-order aspherical mirror (102) and the third freeform surface mirror (103) in sequence and then enters the detector (2). The detector (2) transmits the detection result to the image reconstruction system (3) for processing and outputs an image with near-diffraction-limited image quality. The image restoration algorithm model (301) is a deep learning model based on a physical driving model. The physical driving model is a Wiener filtering algorithm model based on the point spread function of a reflective optical system. The deep learning model is a dynamic convolution deblurring model based on the field attention mechanism. The Wiener filtering algorithm model based on the point spread function of the reflective optical system performs filtering on the input image and the corresponding point spread function PSFs, and adds regularization parameters to finally obtain the deconvolution image. The input to the dynamic convolutional deblurring model based on the field of view attention mechanism includes the field of view coordinate matrix and the deconvolutioned image; The dynamic convolutional deblurring model based on the field of view attention mechanism includes an interconnected FOV encoder and a deformable decoder.

2. The reflective optical system based on computational imaging-driven simplified processing according to claim 1, characterized in that, The F-number of the reflective optical module (1) is in the range of 1 to 14, and the entrance pupil diameter is in the range of 15 mm to 600 mm.

3. The reflective optical system based on computational imaging-driven simplified processing according to claim 1, characterized in that, The surface shapes of the first freeform surface mirror (101), the second even-order aspherical surface mirror (102), and the third freeform surface mirror (103) all satisfy the condition that the root mean square is better than λ / 2, where λ is the wavelength of the light wave.

4. A reflective optical system based on computational imaging-driven simplified processing according to claim 1, characterized in that, The detector (2) is a visible light CMOS detector.

5. A reflective optical system based on computational imaging-driven simplified processing according to claim 1, characterized in that, The FOV encoder includes an FOV module, a first encoding module, a second encoding module, a third encoding module, and a fourth encoding module connected in sequence. The first encoding module includes a convolutional layer Conv and a ResBlock connected in sequence. The second, third, and fourth encoding modules each include a convolutional layer Conv and a ResBlock with a stride of 2 connected in sequence. The deformable decoder includes a first decoding module, a second decoding module, a third decoding module, a fourth decoding module, a KPN module, a first connection module, a second connection module, a third connection module, and a fourth connection module; The first decoding module includes a deformable ResBlock; the second and third decoding modules each include a deconvolutional layer (Deconv), a convolutional layer (Conv), and a deformable ResBlock connected in sequence; the fourth decoding module includes a deconvolutional layer (Deconv), a convolutional layer (Conv), a deformable ResBlock, and a convolutional layer (Conv) connected in sequence; the first connecting module includes an offset transfer layer; the second, third, and fourth connecting modules each include a bilinear upsampling layer (Bilinear Upsampling) and an offset transfer layer (Offset Transfer) connected in sequence. The fourth encoding module, the first decoding module, the second decoding module, the third decoding module, the fourth decoding module, and the KPN module are connected in sequence. The input of the first connecting module is connected to the ResBlock of the fourth encoding module, and the output is used as the input of the first decoding module. In the second connection module, the input of the bilinear upsampling layer is the output of the first connection module, the input of the offset transfer layer also includes the output of the convolutional layer Conv of the second decoding module, the output of the offset transfer layer is connected to the deformable ResBlock of the second decoding module; and the input of the convolutional layer Conv of the second decoding module also includes the output of the ResBlock of the third encoding module. The input of the bilinear upsampling layer in the third connecting module is the output of the offset transfer layer of the second connecting module. The input of the offset transfer layer also includes the output of the convolutional layer Conv of the third decoding module. The output of the offset transfer layer is connected to the deformable ResBlock of the third decoding module. Furthermore, the input of the convolutional layer Conv of the third decoding module also includes the output of the ResBlock of the second encoding module. The input of the Bilinear Upsampling layer in the fourth connecting module is the output of the Offset Transfer layer of the third connecting module and the output of the first convolutional layer Conv of the fourth decoding module. The output of the Offset Transfer layer is connected to the deformable ResBlock of the fourth decoding module; and the input of the convolutional layer Conv of the fourth decoding module also includes the output of the ResBlock of the first encoding module. The input of the KPN module is also connected to the input deconvolution image, and after being superimposed with the input feature map, the images are summarized and processed through a convolution operation to obtain the final output.

6. A reflective optical system based on computational imaging-driven simplified processing according to claim 5, characterized in that, The FOV module first extracts features from the input field of view coordinate matrix FOV through convolutional layers and adds a sigmoid activation function to obtain the FOV feature map; then it processes the input deconvolution image through convolutional layers to obtain the deconvolution feature map; finally, it multiplies the FOV feature map and the deconvolution feature map and then superimposes the result to obtain the output. The deformable ResBlock extracts features based on the offset output by the corresponding connecting module and the input convolutional layer Conv signal. After convolution and superposition with the input convolutional layer Conv signal, the output is obtained.

7. A reflective optical system based on computational imaging-driven simplified processing according to claim 1, characterized in that, The NPU image data processor (302) is an AI chip with the ability to deploy neural networks at the edge. It is used to deploy image restoration algorithm models (301) and meet the computing power required for video-level real-time inference.

Images