Wavefront image processing method based on lightweight neural network, storage medium and equipment
Through the wavefront image processing method based on lightweight neural network, the threshold of each independent sub-aperture in the wavefront image is automatically calculated in real time, solving the problem that the detection accuracy of wavefront sensors in the prior art is affected by environmental factors, and achieving higher detection accuracy and speed.
Patent Information
- Application Number
- CN202411661950.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-11-20
- Publication Date
- 2025-03-14
AI Technical Summary
The existing wavefront sensors are affected by observation environment factors such as temperature, humidity and background noise in terms of detection accuracy, and traditional thresholding methods are difficult to adjust in real time and cannot adapt to the rapidly changing observation environment.
The wavefront image processing method based on lightweight neural network is adopted. The wavefront image is collected by adjusting the wavefront sensor parameters, the threshold of all subapertures on each image is calculated, and the lightweight neural network is constructed for training, and the threshold of each independent subaperture in the wavefront image is automatically calculated in real time.
It improves the detection accuracy and speed of wavefront images, can adapt to changes in the observation environment in real time, and automatically adjusts the threshold, solving the problem that traditional methods cannot adjust and adapt to rapidly changing environments in real time.
Smart Images
Figure CN119625059A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of wavefront image processing, and in particular to a wavefront image processing method, storage medium and device based on a lightweight neural network. Background Art
[0002] As the exploration of the universe deepens, direct imaging of targets such as exoplanets is an important goal in the field of astronomy. However, since the light signals emitted by most observation targets are relatively weak at night, it is very difficult to observe high-value targets such as exoplanets at night. In addition, when the light signal passes through the Earth's atmosphere, it is affected by adverse factors such as air flow disturbances and temperature changes in the Earth's atmosphere, which will cause distortion, resulting in unclear observed targets and the inability to analyze their original physical properties. Therefore, adaptive optical systems came into being to correct the distortion of light signals and allow telescopes to obtain imaging with diffraction-limited capabilities.
[0003] The adaptive optical system consists of a deformable mirror, a wavefront sensor, a controller, and an optical imaging component. The wavefront sensor detects the wavefront slope to calculate the distortion information of the wavefront, and then sends the distortion information to the controller. The controller calculates the control signal according to the distortion information and sends it to the deformable mirror actuator to compensate for the wavefront distortion caused by factors such as atmospheric turbulence by changing the mirror surface shape of the deformable mirror. In order to improve the correction capability of the entire adaptive system, the detection accuracy of the wavefront sensor is crucial. However, the detection accuracy of the wavefront sensor will also be greatly affected by the actual observation environment, such as the temperature, humidity, background noise, etc. of the observation point.
[0004] At present, in order to eliminate the influence of factors such as background noise, most existing wavefront sensors use the threshold method to pre-process the wavefront image obtained by the wavefront sensor, that is, use a pre-calibrated uniform threshold or regional threshold to subtract the background of the wavefront image. Although this type of method is simple, it is very easy to lose information and cannot be adjusted in real time. Although some neural network algorithms can directly calculate the Zenith polynomial coefficients through the image to enhance the accuracy, their network structure is complex and the calculation speed is too slow to keep up with the operating speed of the adaptive optical system. Summary of the invention
[0005] In response to the problems existing in the prior art, the present invention provides a wavefront image processing method, storage medium and device based on a lightweight neural network, which automatically adjusts the threshold of each aperture on the wavefront image in real time, thereby improving the detection accuracy and speed of the wavefront image.
[0006] To achieve the above technical objectives, the present invention adopts the following technical solution: a wavefront image processing method based on a lightweight neural network, specifically comprising the following steps:
[0007] Step S1, adjusting the parameters of the wavefront sensor, collecting the wavefront images in the actual observation process, and calculating the thresholds of all sub-apertures on each wavefront image;
[0008] Step S2, constructing a lightweight neural network, using the calculated thresholds of all sub-apertures on each wavefront image as training labels, and inputting the collected wavefront images into the lightweight neural network in sequence for training;
[0009] Step S3, repeating step S2 until the mean square error loss function converges, completing the training of the lightweight neural network;
[0010] Step S4: Place the trained lightweight neural network into the control loop of the adaptive optical system, perform adaptive sub-aperture threshold calculation on the real-time wavefront image acquired by the wavefront sensor, and improve the spot positioning accuracy of the wavefront image.
[0011] Furthermore, in step S2, the lightweight neural network includes a first feature extraction module, a second feature extraction module, a third feature extraction module, a first fully connected layer and a second fully connected layer which are connected in series in sequence; the first feature extraction module is used to extract basic features of the wavefront image; the second feature extraction module further extracts deep features based on the extracted basic features of the wavefront image; the third feature extraction module further extracts features related to the sub-aperture and spot distribution based on the extracted deep features; the first fully connected layer maps the features extracted by the third feature extraction module to the feature space and extracts global features; the second fully connected layer performs threshold prediction of the sub-aperture based on the extracted global features.
[0012] Furthermore, the first feature extraction module includes a first convolutional layer, a first nonlinear layer and a first pooling layer connected in sequence, the input size of the first convolutional layer is 480×480×1, the convolution kernel is 10×10, the step size is 1, the padding number is 2, and the number of channels is 32, the input size of the first nonlinear layer is 480×480×32, the input size of the first pooling layer is 480×480×32, the pooling window is 4×4, and the step size is 2.
[0013] Furthermore, the second feature extraction module includes: a second convolutional layer, a second nonlinear layer and a second pooling layer connected in sequence, the input size of the second convolutional layer is 120×120×32, the convolution kernel is 8×8, the step size is 1, the padding number is 1, and the number of channels is 32, the input size of the second nonlinear layer is 120×120×32, the input size of the second pooling layer is 120×120×32, the pooling window is 6×6, and the step size is 2.
[0014] Furthermore, the third feature extraction module includes: a third convolutional layer, a third nonlinear layer and a third pooling layer connected in sequence, the input size of the third convolutional layer is 60×60×32, the convolution kernel is 6×6, the step size is 1, the padding number is 1, and the number of channels is 64, the input size of the third nonlinear layer is 120×120×64, the input size of the third pooling layer is 120×120×64, the pooling window is 4×4, and the step size is 2.
[0015] Furthermore, an input size of the first fully connected layer is 1×57600, and an input size of the second fully connected layer is 1×1024.
[0016] Furthermore, the specific process of inputting the collected wavefront image into the lightweight neural network for training in step 2 is as follows:
[0017] Step 2.1, resize all collected wavefront images to 480×480×1;
[0018] Step 2.2, input a wavefront image with a size of 480×480×1 into the first feature extraction module of the lightweight neural network for downsampling, and extract the basic features of the wavefront image with a size of 480×480×32;
[0019] Step 2.2, input the extracted basic features of the wavefront image into the second feature extraction module for downsampling, and extract deep features with a size of 120×120×32;
[0020] Step 2.3, input the extracted deep-level features into the third feature extraction module for downsampling, and extract features with a size of 120×120×64 related to the sub-aperture and spot distribution;
[0021] Step 2.4, mapping the extracted features related to the sub-aperture and spot distribution to the feature space through the first fully connected layer to extract the global features of the wavefront image;
[0022] Step 2.5, the extracted global features are passed through the second fully connected layer to perform sub-aperture threshold prediction;
[0023] Step 2.6: Calculate the mean square error loss function of the predicted sub-aperture threshold and the sub-aperture threshold on the corresponding wavefront image in the training label.
[0024] Furthermore, the calculation process of the mean square error loss function L is:
[0025]
[0026] Where (i, j) represents the coordinates of the subaperture in the wavefront image, I represents the total number of rows of subapertures in the wavefront image, J represents the total number of columns of subapertures in the wavefront image, and T′ ij represents the prediction threshold of the subaperture with coordinates (i, j) in the wavefront image, T ij Represents the threshold of the subaperture with coordinates (i, j) in the wavefront image.
[0027] Furthermore, the present invention also provides a computer-readable storage medium storing a computer program, wherein the computer program enables a computer to execute the wavefront image processing method based on a lightweight neural network.
[0028] Furthermore, the present invention also provides an electronic device, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the wavefront image processing method based on a lightweight neural network is implemented.
[0029] Compared with the prior art, the present invention has the following beneficial effects: the lightweight neural network adopted in the wavefront image processing method based on the lightweight neural network of the present invention is composed of only three convolutional layers, three nonlinear layers, three pooling layers and two fully connected layers, and the network structure is simple. The threshold of each independent sub-aperture in the wavefront image is automatically calculated in real time by the lightweight neural network, so as to improve the signal accuracy of the wavefront image spot signal and thus enhance the detection accuracy of the wavefront sensor. The method can adapt to the observation environment in real time and make adjustments automatically, so as to solve the defects that the traditional unified threshold method or the regional threshold method is not suitable for sensors with a large number of sub-apertures and cannot be automatically adjusted in real time. In addition, the lightweight neural network adopted by the present invention has a simple structure and can run smoothly even on a commercial PC, and can be incorporated into the existing adaptive optical system at a low cost. BRIEF DESCRIPTION OF THE DRAWINGS
[0030] Figure 1 It is a flow chart of the wavefront image processing method based on lightweight neural network of the present invention;
[0031] Figure 2 It is a schematic diagram of the structure of the lightweight neural network in the present invention;
[0032] Figure 3 The wavefront image comparison diagram before and after being processed by the wavefront image processing method based on lightweight neural network of the present invention, wherein: Figure 3 (a) is a schematic diagram of the wavefront image before processing. Figure 3 (b) is a schematic diagram of the wavefront image after processing. DETAILED DESCRIPTION
[0033] The technical solution of the present invention is further explained below in conjunction with the accompanying drawings.
[0034] like Figure 1 The flowchart of the wavefront image processing method based on the lightweight neural network of the present invention comprises the following steps:
[0035] Step S1, adjust the parameters of the wavefront sensor, collect the wavefront images in the actual observation process, and calculate the thresholds of all sub-apertures on each wavefront image; specifically, in order to consider the various situations encountered by the wavefront image in the actual observation process, it is necessary to adjust the inherent parameters of the wavefront sensor, including: input pupil, wavelength, transmittance, sampling method to determine the size and resolution of the wavefront image, among which, in order to obtain the light intensity signal of the input pupil, it is necessary to calculate the random phase of the input pupil
[0036]
[0037] Where (x0, y0) represents the coordinates of the pixel point on the input pupil, K represents the maximum order of the Zenith polynomial corresponding to the input pupil, k represents the index of K, and C i (k) represents the vector coefficient of each order of Zenith polynomial, D k (x0, y0) represents the state equation of the kth order Zennike polynomial of (x0, y0). From this, we can get the light intensity signal U of the input pupil. i (x0, y0):
[0038]
[0039] Among them, A(x0, y0) represents the ideal strongest light signal under any other field conditions where (x0, y0) obeys Gaussian distribution. According to Fourier optics measurement, the calculation formula for the final wavefront image is as follows:
[0040]
[0041] Where f represents the focal length of the microlens array in the wavefront sensor, λ represents the wavelength of the microlens array in the wavefront sensor, T(x0, y0) represents the transmittance of the microlens array in the wavefront sensor, S( ) represents the sampling mode of the wavefront sensor, m represents the lateral sampling range of the wavefront sensor, n represents the longitudinal sampling range of the wavefront sensor, and x f represents the lateral reference coordinate of the wavefront sensor, y f represents the longitudinal reference coordinate of the wavefront sensor, P( ) represents the input pupil equation, and in order to simulate the actual observation environment, random background noise needs to be added to the wavefront image. Then the threshold value of each sub-aperture on the wavefront image is calculated using the following formula:
[0042] T ij =aij +αb ij
[0043] Among them, T ij represents the threshold of the subaperture with coordinates (i, j) in the wavefront image, a ij represents the average noise of the subaperture with coordinates (i, j) in the wavefront image, b ij It represents the noise standard deviation of the sub-aperture with coordinates (i, j) in the wavefront image. α is the proportional coefficient, and its value range is [0.5, 0.8].
[0044] Step S2, construct a lightweight neural network, use the calculated thresholds of all sub-apertures on each wavefront image as training labels, and input the collected wavefront images into the lightweight neural network in sequence for training. The present invention uses a lightweight neural network to automatically calculate the threshold of each independent sub-aperture in the wavefront image in real time, improve the signal accuracy of the wavefront image spot signal, and thus enhance the detection accuracy of the wavefront sensor. It can adapt to the observation environment in real time and make automatic adjustments, thereby solving the problem that the traditional unified threshold method or regional threshold method is not suitable for sensors with a large number of sub-apertures and cannot be automatically adjusted in real time.
[0045] like Figure 2In the present invention, the lightweight neural network includes a first feature extraction module, a second feature extraction module, a third feature extraction module, a first fully connected layer and a second fully connected layer connected in series in sequence; the first feature extraction module is used to extract the basic features of the wavefront image. Specifically, the first feature extraction module includes a first convolutional layer, a first nonlinear layer and a first pooling layer connected in sequence, the input size of the first convolutional layer is 480×480×1, the convolution kernel is 10×10, the step size is 1, the padding number is 2, and the number of channels is 32, the input size of the first nonlinear layer is 480×480×32, the input size of the first pooling layer is 480×480×32, the pooling window is 4×4, and the step size is 2; the second feature extraction module further extracts deep-level features according to the extracted basic features of the wavefront image. Specifically, the second feature extraction module includes: a second convolutional layer, a second nonlinear layer and a second pooling layer connected in sequence, the input size of the second convolutional layer is 120×120×32, the convolution kernel is 8×8, the step size is 1, the padding number is 1, and the number of channels is 32 , the input size of the second nonlinear layer is 120×120×32, the input size of the second pooling layer is 120×120×32, the pooling window is 6×6, and the step size is 2; the third feature extraction module further extracts features related to the sub-aperture and spot distribution according to the extracted deep-level features; the first fully connected layer maps the features extracted by the third feature extraction module to the feature space to extract global features. Specifically, the third feature extraction module includes: a third convolutional layer, a third nonlinear layer and a third pooling layer connected in sequence, and the third The input size of the three convolutional layers is 60×60×32, the convolution kernel is 6×6, the step size is 1, the padding number is 1, and the number of channels is 64. The input size of the third nonlinear layer is 120×120×64, and the input size of the third pooling layer is 120×120×64, the pooling window is 4×4, and the step size is 2. The second fully connected layer predicts the threshold of the sub-aperture according to the extracted global features; specifically, the input size of the first fully connected layer is 1×57600, and the input size of the second fully connected layer is 1×1024. The lightweight neural network consists of only three convolutional layers, three nonlinear layers, three pooling layers, and two fully connected layers. The network structure is simple, and the threshold of each independent sub-aperture in the wavefront image is automatically calculated in real time through the lightweight neural network, which improves the signal accuracy of the wavefront image spot signal and thus enhances the detection accuracy of the wavefront sensor.
[0046] The specific process of inputting the collected wavefront images into the lightweight neural network for training in the present invention is as follows:
[0047] Step 2.1, resize all collected wavefront images to 480×480×1;
[0048] Step 2.2, input a wavefront image with a size of 480×480×1 into the first feature extraction module of the lightweight neural network for downsampling, and extract the basic features of the wavefront image with a size of 480×480×32;
[0049] Step 2.2, input the extracted basic features of the wavefront image into the second feature extraction module for downsampling, and extract deep features with a size of 120×120×32;
[0050] Step 2.3, input the extracted deep-level features into the third feature extraction module for downsampling, and extract features with a size of 120×120×64 related to the sub-aperture and spot distribution;
[0051] Step 2.4, map the extracted features related to the sub-aperture and spot distribution to the feature space through the first fully connected layer, and extract the global features of the wavefront image with a size of 1×57600;
[0052] Step 2.5, the extracted global features are used to predict the sub-aperture threshold through the second fully connected layer with an input size of 1×1024, and 1024 sub-aperture thresholds are predicted for the wavefront image. However, due to the different configurations of the microlens arrays used in the wavefront sensor, there will be different numbers of sub-aperture thresholds. At the same time, the aperture of the telescope system is limited. The usual viewfinder is circular, so for the square microlens, the circular viewfinder will lose some sub-aperture threshold signals. In the method of the present invention, a square microlens array with a configuration of 29×29 is used. The circular viewfinder can only detect 29 sub-aperture thresholds at the full diameter in the middle, so some sub-aperture thresholds will be outside the viewfinder. Therefore, the final number of effective sub-aperture thresholds is only 762;
[0053] Step 2.6: Calculate the mean square error loss function L by comparing the predicted sub-aperture threshold with the threshold of the sub-aperture on the corresponding wavefront image in the training label. The specific calculation process is as follows:
[0054]
[0055] Where (i, j) represents the coordinates of the subaperture in the wavefront image, I represents the total number of rows of the subaperture in the wavefront image, J represents the total number of columns of the subaperture in the wavefront image, and T ij represents the prediction threshold of the subaperture with coordinates (i, j) in the wavefront image, T ij Represents the threshold of the subaperture with coordinates (i, j) in the wavefront image.
[0056] Step S3, repeating step S2 until the mean square error loss function converges, completing the training of the lightweight neural network;
[0057] Step S4: Place the trained lightweight neural network into the control loop of the adaptive optical system, perform adaptive sub-aperture threshold calculation on the real-time wavefront image acquired by the wavefront sensor, and improve the spot positioning accuracy of the wavefront image.
[0058] In one technical solution of the present invention, during the training stage of the lightweight neural network, 7,000 wavefront images are used as a training set, 2,000 wavefront images are used as a test set, and 1,000 wavefront images are used as a verification set. The lightweight neural network is trained according to the training method of the present invention, and the trained lightweight neural network is incorporated into the control loop of the adaptive optical system, that is, the image sent by the wavefront sensor is received, the threshold is automatically calculated in real time, and then sent to the controller to calculate the control command, thereby improving the spot positioning accuracy of the wavefront image. Compared with the existing threshold method, the lightweight neural network only calculates the threshold for network learning during the training stage. After arriving at the observation site and starting to run, the threshold that meets the observation site can be automatically obtained. On-site calibration calculation is no longer required, and it is more suitable for wavefront sensors with a large sub-aperture number. Figure 3 The wavefront image comparison diagram before and after being processed by the wavefront image processing method based on lightweight neural network of the present invention, wherein: Figure 3 (a) is a schematic diagram of the wavefront image before processing. Figure 3 (b) is a schematic diagram of the processed wavefront image. It can be seen that the wavefront image using the wavefront image processing method based on the lightweight neural network of the present invention is clearer than before processing, and the detection accuracy is high, thereby improving the calculation accuracy of the control voltage, making the mirror surface shape change of the deformable mirror most efficient, thereby concentrating the corrected optical path energy to achieve direct imaging of night targets. In addition, most of the existing neural networks used for wavefront sensors are used to directly derive the Zenith polynomial coefficients or control voltages of the wavefront image. Although the wavefront detection part can be skipped, the requirements for hardware equipment are too high and the running time is too long. The lightweight neural network in the wavefront image processing method based on the lightweight neural network of the present invention has a simple structure and can be directly run on a commercial computer without the need for professional equipment such as GPU and DSP.
[0059] In a technical solution of the present invention, a computer-readable storage medium is also provided, storing a computer program, wherein the computer program enables a computer to execute the wavefront image processing method based on a lightweight neural network.
[0060] In a technical solution of the present invention, an electronic device is also provided, including: a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the wavefront image processing method based on a lightweight neural network is implemented.
[0061] In the embodiments disclosed in the present application, the computer storage medium can be a tangible medium that can contain or store a program for use by or in conjunction with an instruction execution system, device or equipment. The computer storage medium can include, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, devices or equipment, or any suitable combination of the above. More specific examples of computer storage media can include electrical connections based on one or more lines, portable computer disks, hard disks, random access memories (RAM), read-only memories (ROM), erasable programmable read-only memories (EPROM or flash memory), optical fibers, portable compact disk read-only memories (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the above.
[0062] Those of ordinary skill in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed in this application can be implemented in electronic hardware or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application.
[0063] The above are only preferred embodiments of the present invention, and the protection scope of the present invention is not limited to the above embodiments. All technical solutions under the concept of the present invention belong to the protection scope of the present invention. It should be pointed out that for ordinary technicians in this technical field, some improvements and modifications without departing from the principle of the present invention should be regarded as the protection scope of the present invention.
Claims
1. A wavefront image processing method based on a lightweight neural network, characterized in that: The specific steps include: Step S1, adjusting the parameters of the wavefront sensor, collecting the wavefront images in the actual observation process, and calculating the thresholds of all sub-apertures on each wavefront image; Step S2, constructing a lightweight neural network, using the calculated thresholds of all sub-apertures on each wavefront image as training labels, and inputting the collected wavefront images into the lightweight neural network in sequence for training; Step S3, repeating step S2 until the mean square error loss function converges, completing the training of the lightweight neural network; Step S4: Place the trained lightweight neural network into the control loop of the adaptive optical system, perform adaptive sub-aperture threshold calculation on the real-time wavefront image acquired by the wavefront sensor, and improve the spot positioning accuracy of the wavefront image.
2. The wavefront image processing method based on lightweight neural network according to claim 1, characterized in that: In step S2, the lightweight neural network includes a first feature extraction module, a second feature extraction module, a third feature extraction module, a first fully connected layer, and a second fully connected layer, which are connected in series in sequence; the first feature extraction module is used to extract basic features of the wavefront image; the second feature extraction module further extracts deep-level features based on the extracted basic features of the wavefront image; the third feature extraction module further extracts features related to sub-aperture and spot distribution based on the extracted deep-level features; the first fully connected layer maps the features extracted by the third feature extraction module to the feature space and extracts global features; the second fully connected layer performs threshold prediction of the sub-aperture based on the extracted global features.
3. The wavefront image processing method based on lightweight neural network according to claim 2 is characterized in that: The first feature extraction module includes a first convolutional layer, a first nonlinear layer and a first pooling layer connected in sequence, the input size of the first convolutional layer is 480×480×1, the convolution kernel is 10×10, the step size is 1, the padding number is 2, and the number of channels is 32, the input size of the first nonlinear layer is 480×480×32, the input size of the first pooling layer is 480×480×32, the pooling window is 4×4, and the step size is 2.
4. The wavefront image processing method based on lightweight neural network according to claim 3 is characterized in that: The second feature extraction module includes: a second convolutional layer, a second nonlinear layer and a second pooling layer connected in sequence, the input size of the second convolutional layer is 120×120×32, the convolution kernel is 8×8, the step size is 1, the padding number is 1, and the number of channels is 32, the input size of the second nonlinear layer is 120×120×32, the input size of the second pooling layer is 120×120×32, the pooling window is 6×6, and the step size is 2.
5. The wavefront image processing method based on lightweight neural network according to claim 4 is characterized in that: The third feature extraction module includes: a third convolutional layer, a third nonlinear layer and a third pooling layer connected in sequence, the input size of the third convolutional layer is 60×60×32, the convolution kernel is 6×6, the step size is 1, the padding number is 1, and the number of channels is 64, the input size of the third nonlinear layer is 120×120×64, the input size of the third pooling layer is 120×120×64, the pooling window is 4×4, and the step size is 2.
6. The wavefront image processing method based on lightweight neural network according to claim 5, characterized in that: The input size of the first fully connected layer is 1×57600, and the input size of the second fully connected layer is 1×1024.
7. The wavefront image processing method based on lightweight neural network according to claim 6, characterized in that: The specific process of inputting the collected wavefront image into the lightweight neural network for training in step 2 is as follows: Step 2.1, resize all collected wavefront images to 480×480×1; Step 2.2, input a wavefront image with a size of 480×480×1 into the first feature extraction module of the lightweight neural network for downsampling, and extract the basic features of the wavefront image with a size of 480×480×32; Step 2.2, input the extracted basic features of the wavefront image into the second feature extraction module for downsampling, and extract deep features with a size of 120×120×32; Step 2.3, input the extracted deep-level features into the third feature extraction module for downsampling, and extract features with a size of 120×120×64 related to the sub-aperture and spot distribution; Step 2.4, mapping the extracted features related to the sub-aperture and spot distribution to the feature space through the first fully connected layer to extract the global features of the wavefront image; Step 2.5, the extracted global features are passed through the second fully connected layer to perform sub-aperture threshold prediction; Step 2.6: Calculate the mean square error loss function of the predicted sub-aperture threshold and the sub-aperture threshold on the corresponding wavefront image in the training label.
8. The wavefront image processing method based on lightweight neural network according to claim 7, characterized in that: The calculation process of the mean square error loss function L is: Where (i, j) represents the coordinates of the subaperture in the wavefront image, I represents the total number of rows of subapertures in the wavefront image, J represents the total number of columns of subapertures in the wavefront image, and T' ij represents the prediction threshold of the subaperture with coordinates (i, j) in the wavefront image, T ij Represents the threshold of the subaperture with coordinates (i, j) in the wavefront image.
9. A computer-readable storage medium storing a computer program, characterized in that: The computer program enables the computer to execute a wavefront image processing method based on a lightweight neural network as described in any one of claims 1 to 8.
10. An electronic device, characterized in that: include: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, a wavefront image processing method based on a lightweight neural network as described in any one of claims 1 to 8 is implemented.